Gitiles
Code Review Sign In
gerrit-public.fairphone.software / fp2-dev / platform / external / llvm / 845b4c2f0b2b3f0ec468ac863550b25708279180 / . / lib / Target / X86 / X86InstrInfo.cpp
blob: 7899985badc0d8eb777529a578cf2b1b6dbf6e7f [file] [log] [blame]
Jia Liu31d157a2012-02-18 12:03:15 +0000[diff] [blame]1//===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
Misha Brukman0e0a7a452005-04-21 23:38:14 +0000[diff] [blame]2//
John Criswellb576c942003-10-20 19:43:21 +0000[diff] [blame]3// The LLVM Compiler Infrastructure
4//
Chris Lattner4ee451d2007-12-29 20:36:04 +0000[diff] [blame]5// 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 +0000[diff] [blame]7//
John Criswellb576c942003-10-20 19:43:21 +0000[diff] [blame]8//===----------------------------------------------------------------------===//
Chris Lattner72614082002-10-25 22:55:53 +0000[diff] [blame]9//
Chris Lattner3501fea2003-01-14 22:00:31 +0000[diff] [blame]10// This file contains the X86 implementation of the TargetInstrInfo class.
Chris Lattner72614082002-10-25 22:55:53 +0000[diff] [blame]11//
12//===----------------------------------------------------------------------===//
13
Chris Lattner055c9652002-10-29 21:05:24 +0000[diff] [blame]14#include "X86InstrInfo.h"
Chris Lattner4ce42a72002-12-03 05:42:53 +0000[diff] [blame]15#include "X86.h"
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]16#include "X86InstrBuilder.h"
Owen Andersond94b6a12008-01-04 23:57:37 +0000[diff] [blame]17#include "X86MachineFunctionInfo.h"
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]18#include "X86Subtarget.h"
19#include "X86TargetMachine.h"
Dan Gohmand68a0762009-01-05 17:59:02 +0000[diff] [blame]20#include "llvm/DerivedTypes.h"
Owen Anderson0a5372e2009-07-13 04:09:18 +0000[diff] [blame]21#include "llvm/LLVMContext.h"
Owen Anderson718cb662007-09-07 04:06:50 +0000[diff] [blame]22#include "llvm/ADT/STLExtras.h"
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]23#include "llvm/CodeGen/MachineConstantPool.h"
Hans Wennborgf0234fc2012-06-01 16:27:21 +0000[diff] [blame]24#include "llvm/CodeGen/MachineDominators.h"
Owen Andersond94b6a12008-01-04 23:57:37 +0000[diff] [blame]25#include "llvm/CodeGen/MachineFrameInfo.h"
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]26#include "llvm/CodeGen/MachineInstrBuilder.h"
Chris Lattner84bc5422007-12-31 04:13:23 +0000[diff] [blame]27#include "llvm/CodeGen/MachineRegisterInfo.h"
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]28#include "llvm/CodeGen/LiveVariables.h"
Craig Topper79aa3412012-03-17 18:46:09 +0000[diff] [blame]29#include "llvm/MC/MCAsmInfo.h"
Chris Lattneree9eb412010-04-26 23:37:21 +0000[diff] [blame]30#include "llvm/MC/MCInst.h"
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]31#include "llvm/Support/CommandLine.h"
David Greene5b901322010-01-05 01:29:29 +0000[diff] [blame]32#include "llvm/Support/Debug.h"
Torok Edwinab7c09b2009-07-08 18:01:40 +0000[diff] [blame]33#include "llvm/Support/ErrorHandling.h"
34#include "llvm/Support/raw_ostream.h"
Evan Cheng0488db92007-09-25 01:57:46 +0000[diff] [blame]35#include "llvm/Target/TargetOptions.h"
David Greeneb87bc952009-11-12 20:55:29 +0000[diff] [blame]36#include <limits>
37
Evan Cheng4db3cff2011-07-01 17:57:27 +0000[diff] [blame]38#define GET_INSTRINFO_CTOR
Evan Cheng22fee2d2011-06-28 20:07:07 +0000[diff] [blame]39#include "X86GenInstrInfo.inc"
40
Brian Gaeked0fde302003-11-11 22:41:34 +0000[diff] [blame]41using namespace llvm;
42
Chris Lattner705e07f2009-08-23 03:41:05 +0000[diff] [blame]43static 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 +0000[diff] [blame]55
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]56enum {
57 // Select which memory operand is being unfolded.
Craig Topper3ed920f2012-06-23 08:01:18 +0000[diff] [blame]58 // (stored in bits 0 - 3)
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]59 TB_INDEX_0 = 0,
60 TB_INDEX_1 = 1,
61 TB_INDEX_2 = 2,
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]62 TB_INDEX_3 = 3,
Craig Topper3ed920f2012-06-23 08:01:18 +0000[diff] [blame]63 TB_INDEX_MASK = 0xf,
64
65 // Do not insert the reverse map (MemOp -> RegOp) into the table.
66 // This may be needed because there is a many -> one mapping.
67 TB_NO_REVERSE = 1 << 4,
68
69 // Do not insert the forward map (RegOp -> MemOp) into the table.
70 // This is needed for Native Client, which prohibits branch
71 // instructions from using a memory operand.
72 TB_NO_FORWARD = 1 << 5,
73
74 TB_FOLDED_LOAD = 1 << 6,
75 TB_FOLDED_STORE = 1 << 7,
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]76
77 // Minimum alignment required for load/store.
78 // Used for RegOp->MemOp conversion.
79 // (stored in bits 8 - 15)
80 TB_ALIGN_SHIFT = 8,
81 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
82 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
83 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
Craig Topper3ed920f2012-06-23 08:01:18 +0000[diff] [blame]84 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]85};
86
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]87struct X86OpTblEntry {
88 uint16_t RegOp;
89 uint16_t MemOp;
Craig Topper3ed920f2012-06-23 08:01:18 +0000[diff] [blame]90 uint16_t Flags;
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]91};
92
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]93X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
Evan Cheng4db3cff2011-07-01 17:57:27 +0000[diff] [blame]94 : 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 +0000[diff] [blame]100 TM(tm), RI(tm, *this) {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]101
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]102 static const X86OpTblEntry OpTbl2Addr[] = {
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]103 { X86::ADC32ri, X86::ADC32mi, 0 },
104 { X86::ADC32ri8, X86::ADC32mi8, 0 },
105 { X86::ADC32rr, X86::ADC32mr, 0 },
106 { X86::ADC64ri32, X86::ADC64mi32, 0 },
107 { X86::ADC64ri8, X86::ADC64mi8, 0 },
108 { X86::ADC64rr, X86::ADC64mr, 0 },
109 { X86::ADD16ri, X86::ADD16mi, 0 },
110 { X86::ADD16ri8, X86::ADD16mi8, 0 },
111 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
112 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
113 { X86::ADD16rr, X86::ADD16mr, 0 },
114 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
115 { X86::ADD32ri, X86::ADD32mi, 0 },
116 { X86::ADD32ri8, X86::ADD32mi8, 0 },
117 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
118 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
119 { X86::ADD32rr, X86::ADD32mr, 0 },
120 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
121 { X86::ADD64ri32, X86::ADD64mi32, 0 },
122 { X86::ADD64ri8, X86::ADD64mi8, 0 },
123 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
124 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
125 { X86::ADD64rr, X86::ADD64mr, 0 },
126 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
127 { X86::ADD8ri, X86::ADD8mi, 0 },
128 { X86::ADD8rr, X86::ADD8mr, 0 },
129 { X86::AND16ri, X86::AND16mi, 0 },
130 { X86::AND16ri8, X86::AND16mi8, 0 },
131 { X86::AND16rr, X86::AND16mr, 0 },
132 { X86::AND32ri, X86::AND32mi, 0 },
133 { X86::AND32ri8, X86::AND32mi8, 0 },
134 { X86::AND32rr, X86::AND32mr, 0 },
135 { X86::AND64ri32, X86::AND64mi32, 0 },
136 { X86::AND64ri8, X86::AND64mi8, 0 },
137 { X86::AND64rr, X86::AND64mr, 0 },
138 { X86::AND8ri, X86::AND8mi, 0 },
139 { X86::AND8rr, X86::AND8mr, 0 },
140 { X86::DEC16r, X86::DEC16m, 0 },
141 { X86::DEC32r, X86::DEC32m, 0 },
142 { X86::DEC64_16r, X86::DEC64_16m, 0 },
143 { X86::DEC64_32r, X86::DEC64_32m, 0 },
144 { X86::DEC64r, X86::DEC64m, 0 },
145 { X86::DEC8r, X86::DEC8m, 0 },
146 { X86::INC16r, X86::INC16m, 0 },
147 { X86::INC32r, X86::INC32m, 0 },
148 { X86::INC64_16r, X86::INC64_16m, 0 },
149 { X86::INC64_32r, X86::INC64_32m, 0 },
150 { X86::INC64r, X86::INC64m, 0 },
151 { X86::INC8r, X86::INC8m, 0 },
152 { X86::NEG16r, X86::NEG16m, 0 },
153 { X86::NEG32r, X86::NEG32m, 0 },
154 { X86::NEG64r, X86::NEG64m, 0 },
155 { X86::NEG8r, X86::NEG8m, 0 },
156 { X86::NOT16r, X86::NOT16m, 0 },
157 { X86::NOT32r, X86::NOT32m, 0 },
158 { X86::NOT64r, X86::NOT64m, 0 },
159 { X86::NOT8r, X86::NOT8m, 0 },
160 { X86::OR16ri, X86::OR16mi, 0 },
161 { X86::OR16ri8, X86::OR16mi8, 0 },
162 { X86::OR16rr, X86::OR16mr, 0 },
163 { X86::OR32ri, X86::OR32mi, 0 },
164 { X86::OR32ri8, X86::OR32mi8, 0 },
165 { X86::OR32rr, X86::OR32mr, 0 },
166 { X86::OR64ri32, X86::OR64mi32, 0 },
167 { X86::OR64ri8, X86::OR64mi8, 0 },
168 { X86::OR64rr, X86::OR64mr, 0 },
169 { X86::OR8ri, X86::OR8mi, 0 },
170 { X86::OR8rr, X86::OR8mr, 0 },
171 { X86::ROL16r1, X86::ROL16m1, 0 },
172 { X86::ROL16rCL, X86::ROL16mCL, 0 },
173 { X86::ROL16ri, X86::ROL16mi, 0 },
174 { X86::ROL32r1, X86::ROL32m1, 0 },
175 { X86::ROL32rCL, X86::ROL32mCL, 0 },
176 { X86::ROL32ri, X86::ROL32mi, 0 },
177 { X86::ROL64r1, X86::ROL64m1, 0 },
178 { X86::ROL64rCL, X86::ROL64mCL, 0 },
179 { X86::ROL64ri, X86::ROL64mi, 0 },
180 { X86::ROL8r1, X86::ROL8m1, 0 },
181 { X86::ROL8rCL, X86::ROL8mCL, 0 },
182 { X86::ROL8ri, X86::ROL8mi, 0 },
183 { X86::ROR16r1, X86::ROR16m1, 0 },
184 { X86::ROR16rCL, X86::ROR16mCL, 0 },
185 { X86::ROR16ri, X86::ROR16mi, 0 },
186 { X86::ROR32r1, X86::ROR32m1, 0 },
187 { X86::ROR32rCL, X86::ROR32mCL, 0 },
188 { X86::ROR32ri, X86::ROR32mi, 0 },
189 { X86::ROR64r1, X86::ROR64m1, 0 },
190 { X86::ROR64rCL, X86::ROR64mCL, 0 },
191 { X86::ROR64ri, X86::ROR64mi, 0 },
192 { X86::ROR8r1, X86::ROR8m1, 0 },
193 { X86::ROR8rCL, X86::ROR8mCL, 0 },
194 { X86::ROR8ri, X86::ROR8mi, 0 },
195 { X86::SAR16r1, X86::SAR16m1, 0 },
196 { X86::SAR16rCL, X86::SAR16mCL, 0 },
197 { X86::SAR16ri, X86::SAR16mi, 0 },
198 { X86::SAR32r1, X86::SAR32m1, 0 },
199 { X86::SAR32rCL, X86::SAR32mCL, 0 },
200 { X86::SAR32ri, X86::SAR32mi, 0 },
201 { X86::SAR64r1, X86::SAR64m1, 0 },
202 { X86::SAR64rCL, X86::SAR64mCL, 0 },
203 { X86::SAR64ri, X86::SAR64mi, 0 },
204 { X86::SAR8r1, X86::SAR8m1, 0 },
205 { X86::SAR8rCL, X86::SAR8mCL, 0 },
206 { X86::SAR8ri, X86::SAR8mi, 0 },
207 { X86::SBB32ri, X86::SBB32mi, 0 },
208 { X86::SBB32ri8, X86::SBB32mi8, 0 },
209 { X86::SBB32rr, X86::SBB32mr, 0 },
210 { X86::SBB64ri32, X86::SBB64mi32, 0 },
211 { X86::SBB64ri8, X86::SBB64mi8, 0 },
212 { X86::SBB64rr, X86::SBB64mr, 0 },
213 { X86::SHL16rCL, X86::SHL16mCL, 0 },
214 { X86::SHL16ri, X86::SHL16mi, 0 },
215 { X86::SHL32rCL, X86::SHL32mCL, 0 },
216 { X86::SHL32ri, X86::SHL32mi, 0 },
217 { X86::SHL64rCL, X86::SHL64mCL, 0 },
218 { X86::SHL64ri, X86::SHL64mi, 0 },
219 { X86::SHL8rCL, X86::SHL8mCL, 0 },
220 { X86::SHL8ri, X86::SHL8mi, 0 },
221 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
222 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
223 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
224 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
225 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
226 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
227 { X86::SHR16r1, X86::SHR16m1, 0 },
228 { X86::SHR16rCL, X86::SHR16mCL, 0 },
229 { X86::SHR16ri, X86::SHR16mi, 0 },
230 { X86::SHR32r1, X86::SHR32m1, 0 },
231 { X86::SHR32rCL, X86::SHR32mCL, 0 },
232 { X86::SHR32ri, X86::SHR32mi, 0 },
233 { X86::SHR64r1, X86::SHR64m1, 0 },
234 { X86::SHR64rCL, X86::SHR64mCL, 0 },
235 { X86::SHR64ri, X86::SHR64mi, 0 },
236 { X86::SHR8r1, X86::SHR8m1, 0 },
237 { X86::SHR8rCL, X86::SHR8mCL, 0 },
238 { X86::SHR8ri, X86::SHR8mi, 0 },
239 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
240 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
241 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
242 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
243 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
244 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
245 { X86::SUB16ri, X86::SUB16mi, 0 },
246 { X86::SUB16ri8, X86::SUB16mi8, 0 },
247 { X86::SUB16rr, X86::SUB16mr, 0 },
248 { X86::SUB32ri, X86::SUB32mi, 0 },
249 { X86::SUB32ri8, X86::SUB32mi8, 0 },
250 { X86::SUB32rr, X86::SUB32mr, 0 },
251 { X86::SUB64ri32, X86::SUB64mi32, 0 },
252 { X86::SUB64ri8, X86::SUB64mi8, 0 },
253 { X86::SUB64rr, X86::SUB64mr, 0 },
254 { X86::SUB8ri, X86::SUB8mi, 0 },
255 { X86::SUB8rr, X86::SUB8mr, 0 },
256 { X86::XOR16ri, X86::XOR16mi, 0 },
257 { X86::XOR16ri8, X86::XOR16mi8, 0 },
258 { X86::XOR16rr, X86::XOR16mr, 0 },
259 { X86::XOR32ri, X86::XOR32mi, 0 },
260 { X86::XOR32ri8, X86::XOR32mi8, 0 },
261 { X86::XOR32rr, X86::XOR32mr, 0 },
262 { X86::XOR64ri32, X86::XOR64mi32, 0 },
263 { X86::XOR64ri8, X86::XOR64mi8, 0 },
264 { X86::XOR64rr, X86::XOR64mr, 0 },
265 { X86::XOR8ri, X86::XOR8mi, 0 },
266 { X86::XOR8rr, X86::XOR8mr, 0 }
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]267 };
268
269 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]270 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 +0000[diff] [blame]273 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 +0000[diff] [blame]277 }
278
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]279 static const X86OpTblEntry OpTbl0[] = {
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]280 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
281 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
282 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
283 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
284 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]285 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
286 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
287 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
288 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
289 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
290 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
291 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
292 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
293 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
294 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
295 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
296 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
297 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
298 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
299 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
300 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
301 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
302 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]303 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
304 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
305 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
306 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
307 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
308 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
309 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
310 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
311 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
312 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
313 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
314 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
315 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
316 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
317 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
318 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
319 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
320 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
321 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
322 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
323 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
324 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]325 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
326 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
327 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
328 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
329 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
330 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]331 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
332 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
333 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
334 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
335 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
336 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
337 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
338 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
339 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
340 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
341 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
342 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
343 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
344 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
345 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
346 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
347 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
348 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
349 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
350 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
351 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
352 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
353 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
354 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
355 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]356 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
357 // AVX 128-bit versions of foldable instructions
358 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
359 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
360 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]361 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]362 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
363 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
366 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
367 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
368 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
369 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
370 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
371 // AVX 256-bit foldable instructions
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]372 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]373 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
374 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
377 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]378 };
379
380 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]381 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 +0000[diff] [blame]384 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
385 RegOp, MemOp, TB_INDEX_0 | Flags);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]386 }
387
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]388 static const X86OpTblEntry OpTbl1[] = {
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]389 { X86::CMP16rr, X86::CMP16rm, 0 },
390 { X86::CMP32rr, X86::CMP32rm, 0 },
391 { X86::CMP64rr, X86::CMP64rm, 0 },
392 { X86::CMP8rr, X86::CMP8rm, 0 },
393 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
394 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
395 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
396 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
397 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
398 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
399 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
400 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
401 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
402 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
403 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
404 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]405 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
406 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
407 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
408 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
409 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
410 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
411 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
412 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]413 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
414 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
Craig Topper8e58b3e2012-06-15 07:02:58 +0000[diff] [blame]415 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
416 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]417 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
418 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
419 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
420 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
421 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
422 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
423 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
424 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
425 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
426 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
427 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
428 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
429 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
430 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]431 { X86::MOV16rr, X86::MOV16rm, 0 },
432 { X86::MOV32rr, X86::MOV32rm, 0 },
433 { X86::MOV64rr, X86::MOV64rm, 0 },
434 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
435 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
436 { X86::MOV8rr, X86::MOV8rm, 0 },
437 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
438 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]439 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
440 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
441 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
442 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]443 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
444 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
445 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
446 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
447 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
448 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
449 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
450 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
451 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
452 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]453 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
454 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
455 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
456 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
457 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
458 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
459 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
460 { X86::MOVZX64rr16, X86::MOVZX64rm16, 0 },
461 { X86::MOVZX64rr32, X86::MOVZX64rm32, 0 },
462 { X86::MOVZX64rr8, X86::MOVZX64rm8, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]463 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
464 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
465 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]466 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
467 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
468 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
469 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
470 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
471 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
472 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
473 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
474 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
475 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
476 { X86::SQRTPDr_Int, X86::SQRTPDm_Int, TB_ALIGN_16 },
477 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
478 { X86::SQRTPSr_Int, X86::SQRTPSm_Int, TB_ALIGN_16 },
479 { X86::SQRTSDr, X86::SQRTSDm, 0 },
480 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
481 { X86::SQRTSSr, X86::SQRTSSm, 0 },
482 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
483 { X86::TEST16rr, X86::TEST16rm, 0 },
484 { X86::TEST32rr, X86::TEST32rm, 0 },
485 { X86::TEST64rr, X86::TEST64rm, 0 },
486 { X86::TEST8rr, X86::TEST8rm, 0 },
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]487 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]488 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
489 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]490 // AVX 128-bit versions of foldable instructions
491 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
492 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]493 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
494 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
Craig Topper8e58b3e2012-06-15 07:02:58 +0000[diff] [blame]495 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
496 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
Pete Cooper312091e2012-06-14 22:12:58 +0000[diff] [blame]497 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
Craig Topper8e58b3e2012-06-15 07:02:58 +0000[diff] [blame]498 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
499 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
500 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
501 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
502 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
503 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
504 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
505 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
506 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]507 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
508 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
509 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
510 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
511 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
512 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
513 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
514 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
515 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
516 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
517 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
518 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
519 { X86::VMOVUPDrr, X86::VMOVUPDrm, TB_ALIGN_16 },
520 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
521 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
522 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
523 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]524 { X86::VPABSBrr128, X86::VPABSBrm128, TB_ALIGN_16 },
525 { X86::VPABSDrr128, X86::VPABSDrm128, TB_ALIGN_16 },
526 { X86::VPABSWrr128, X86::VPABSWrm128, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]527 { X86::VPERMILPDri, X86::VPERMILPDmi, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]528 { X86::VPERMILPSri, X86::VPERMILPSmi, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]529 { X86::VPSHUFDri, X86::VPSHUFDmi, TB_ALIGN_16 },
530 { X86::VPSHUFHWri, X86::VPSHUFHWmi, TB_ALIGN_16 },
531 { X86::VPSHUFLWri, X86::VPSHUFLWmi, TB_ALIGN_16 },
532 { X86::VRCPPSr, X86::VRCPPSm, TB_ALIGN_16 },
533 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, TB_ALIGN_16 },
534 { X86::VRSQRTPSr, X86::VRSQRTPSm, TB_ALIGN_16 },
535 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, TB_ALIGN_16 },
536 { X86::VSQRTPDr, X86::VSQRTPDm, TB_ALIGN_16 },
537 { X86::VSQRTPDr_Int, X86::VSQRTPDm_Int, TB_ALIGN_16 },
538 { X86::VSQRTPSr, X86::VSQRTPSm, TB_ALIGN_16 },
539 { X86::VSQRTPSr_Int, X86::VSQRTPSm_Int, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]540 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]541 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
542 // AVX 256-bit foldable instructions
543 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
544 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
Craig Topper40385c82012-01-19 08:50:38 +0000[diff] [blame]545 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]546 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]547 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
Craig Topper40385c82012-01-19 08:50:38 +0000[diff] [blame]548 { X86::VPERMILPDYri, X86::VPERMILPDYmi, TB_ALIGN_32 },
549 { X86::VPERMILPSYri, X86::VPERMILPSYmi, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]550 // AVX2 foldable instructions
Craig Topper40385c82012-01-19 08:50:38 +0000[diff] [blame]551 { X86::VPABSBrr256, X86::VPABSBrm256, TB_ALIGN_32 },
552 { X86::VPABSDrr256, X86::VPABSDrm256, TB_ALIGN_32 },
553 { X86::VPABSWrr256, X86::VPABSWrm256, TB_ALIGN_32 },
554 { X86::VPSHUFDYri, X86::VPSHUFDYmi, TB_ALIGN_32 },
555 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, TB_ALIGN_32 },
556 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, TB_ALIGN_32 },
557 { X86::VRCPPSYr, X86::VRCPPSYm, TB_ALIGN_32 },
558 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, TB_ALIGN_32 },
559 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, TB_ALIGN_32 },
560 { X86::VRSQRTPSYr_Int, X86::VRSQRTPSYm_Int, TB_ALIGN_32 },
561 { X86::VSQRTPDYr, X86::VSQRTPDYm, TB_ALIGN_32 },
562 { X86::VSQRTPDYr_Int, X86::VSQRTPDYm_Int, TB_ALIGN_32 },
563 { X86::VSQRTPSYr, X86::VSQRTPSYm, TB_ALIGN_32 },
564 { X86::VSQRTPSYr_Int, X86::VSQRTPSYm_Int, TB_ALIGN_32 },
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]565 };
566
567 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]568 unsigned RegOp = OpTbl1[i].RegOp;
569 unsigned MemOp = OpTbl1[i].MemOp;
570 unsigned Flags = OpTbl1[i].Flags;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]571 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
572 RegOp, MemOp,
573 // Index 1, folded load
574 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]575 }
576
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]577 static const X86OpTblEntry OpTbl2[] = {
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]578 { X86::ADC32rr, X86::ADC32rm, 0 },
579 { X86::ADC64rr, X86::ADC64rm, 0 },
580 { X86::ADD16rr, X86::ADD16rm, 0 },
581 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
582 { X86::ADD32rr, X86::ADD32rm, 0 },
583 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
584 { X86::ADD64rr, X86::ADD64rm, 0 },
585 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
586 { X86::ADD8rr, X86::ADD8rm, 0 },
587 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
588 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
589 { X86::ADDSDrr, X86::ADDSDrm, 0 },
590 { X86::ADDSSrr, X86::ADDSSrm, 0 },
591 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
592 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
593 { X86::AND16rr, X86::AND16rm, 0 },
594 { X86::AND32rr, X86::AND32rm, 0 },
595 { X86::AND64rr, X86::AND64rm, 0 },
596 { X86::AND8rr, X86::AND8rm, 0 },
597 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
598 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
599 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
600 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]601 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
602 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
603 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
604 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]605 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
606 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
607 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
608 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
609 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
610 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
611 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
612 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
613 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
614 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
615 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
616 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
617 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
618 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
619 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
620 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
621 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
622 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
623 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
624 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
625 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
626 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
627 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
628 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
629 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
630 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
631 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
632 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
633 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
634 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
635 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
636 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
637 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
638 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
639 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
640 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
641 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
642 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
643 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
644 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
645 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
646 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
647 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
648 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
649 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
650 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
651 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
652 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
653 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
654 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
655 { X86::CMPSDrr, X86::CMPSDrm, 0 },
656 { X86::CMPSSrr, X86::CMPSSrm, 0 },
657 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
658 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
659 { X86::DIVSDrr, X86::DIVSDrm, 0 },
660 { X86::DIVSSrr, X86::DIVSSrm, 0 },
661 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
662 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
663 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
664 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
665 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
666 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
667 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
668 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
669 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
670 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
671 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
672 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
673 { X86::IMUL16rr, X86::IMUL16rm, 0 },
674 { X86::IMUL32rr, X86::IMUL32rm, 0 },
675 { X86::IMUL64rr, X86::IMUL64rm, 0 },
676 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
677 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
678 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
679 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, TB_ALIGN_16 },
680 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
681 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, TB_ALIGN_16 },
682 { X86::MAXSDrr, X86::MAXSDrm, 0 },
683 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
684 { X86::MAXSSrr, X86::MAXSSrm, 0 },
685 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
686 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
687 { X86::MINPDrr_Int, X86::MINPDrm_Int, TB_ALIGN_16 },
688 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
689 { X86::MINPSrr_Int, X86::MINPSrm_Int, TB_ALIGN_16 },
690 { X86::MINSDrr, X86::MINSDrm, 0 },
691 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
692 { X86::MINSSrr, X86::MINSSrm, 0 },
693 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]694 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]695 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
696 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
697 { X86::MULSDrr, X86::MULSDrm, 0 },
698 { X86::MULSSrr, X86::MULSSrm, 0 },
699 { X86::OR16rr, X86::OR16rm, 0 },
700 { X86::OR32rr, X86::OR32rm, 0 },
701 { X86::OR64rr, X86::OR64rm, 0 },
702 { X86::OR8rr, X86::OR8rm, 0 },
703 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
704 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
705 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
706 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]707 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]708 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
709 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
710 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
711 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
712 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
713 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]714 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
715 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]716 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]717 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]718 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
719 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
720 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
721 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]722 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]723 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
724 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]725 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]726 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
727 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
728 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]729 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]730 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]731 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
732 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]733 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]734 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]735 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]736 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]737 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]738 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]739 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
740 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
741 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
742 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
743 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
744 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]745 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]746 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
747 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
748 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
749 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
750 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
751 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
752 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
Craig Topper969ba282012-01-25 06:43:11 +0000[diff] [blame]753 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
754 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
755 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
756 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]757 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
758 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
759 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
760 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
761 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
762 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
763 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
764 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
765 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
766 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
767 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
768 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
769 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
770 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
771 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
772 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
773 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
774 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
775 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
776 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
777 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
778 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
779 { X86::SBB32rr, X86::SBB32rm, 0 },
780 { X86::SBB64rr, X86::SBB64rm, 0 },
781 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
782 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
783 { X86::SUB16rr, X86::SUB16rm, 0 },
784 { X86::SUB32rr, X86::SUB32rm, 0 },
785 { X86::SUB64rr, X86::SUB64rm, 0 },
786 { X86::SUB8rr, X86::SUB8rm, 0 },
787 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
788 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
789 { X86::SUBSDrr, X86::SUBSDrm, 0 },
790 { X86::SUBSSrr, X86::SUBSSrm, 0 },
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]791 // FIXME: TEST*rr -> swapped operand of TEST*mr.
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]792 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
793 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
794 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
795 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
796 { X86::XOR16rr, X86::XOR16rm, 0 },
797 { X86::XOR32rr, X86::XOR32rm, 0 },
798 { X86::XOR64rr, X86::XOR64rm, 0 },
799 { X86::XOR8rr, X86::XOR8rm, 0 },
800 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]801 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
802 // AVX 128-bit versions of foldable instructions
803 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
804 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
805 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
806 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
807 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
808 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
809 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
810 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
811 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
812 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
813 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
814 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
Craig Topper13d89c72012-06-25 06:16:00 +0000[diff] [blame]815 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]816 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, TB_ALIGN_16 },
817 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
818 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
819 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
820 { X86::VADDPDrr, X86::VADDPDrm, TB_ALIGN_16 },
821 { X86::VADDPSrr, X86::VADDPSrm, TB_ALIGN_16 },
822 { X86::VADDSDrr, X86::VADDSDrm, 0 },
823 { X86::VADDSSrr, X86::VADDSSrm, 0 },
824 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, TB_ALIGN_16 },
825 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, TB_ALIGN_16 },
826 { X86::VANDNPDrr, X86::VANDNPDrm, TB_ALIGN_16 },
827 { X86::VANDNPSrr, X86::VANDNPSrm, TB_ALIGN_16 },
828 { X86::VANDPDrr, X86::VANDPDrm, TB_ALIGN_16 },
829 { X86::VANDPSrr, X86::VANDPSrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]830 { X86::VBLENDPDrri, X86::VBLENDPDrmi, TB_ALIGN_16 },
831 { X86::VBLENDPSrri, X86::VBLENDPSrmi, TB_ALIGN_16 },
832 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, TB_ALIGN_16 },
833 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]834 { X86::VCMPPDrri, X86::VCMPPDrmi, TB_ALIGN_16 },
835 { X86::VCMPPSrri, X86::VCMPPSrmi, TB_ALIGN_16 },
836 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
837 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
838 { X86::VDIVPDrr, X86::VDIVPDrm, TB_ALIGN_16 },
839 { X86::VDIVPSrr, X86::VDIVPSrm, TB_ALIGN_16 },
840 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
841 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
842 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
843 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
844 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
845 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
846 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
847 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
848 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
849 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
850 { X86::VHADDPDrr, X86::VHADDPDrm, TB_ALIGN_16 },
851 { X86::VHADDPSrr, X86::VHADDPSrm, TB_ALIGN_16 },
852 { X86::VHSUBPDrr, X86::VHSUBPDrm, TB_ALIGN_16 },
853 { X86::VHSUBPSrr, X86::VHSUBPSrm, TB_ALIGN_16 },
854 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
855 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
856 { X86::VMAXPDrr, X86::VMAXPDrm, TB_ALIGN_16 },
857 { X86::VMAXPDrr_Int, X86::VMAXPDrm_Int, TB_ALIGN_16 },
858 { X86::VMAXPSrr, X86::VMAXPSrm, TB_ALIGN_16 },
859 { X86::VMAXPSrr_Int, X86::VMAXPSrm_Int, TB_ALIGN_16 },
860 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
861 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 },
862 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
863 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 },
864 { X86::VMINPDrr, X86::VMINPDrm, TB_ALIGN_16 },
865 { X86::VMINPDrr_Int, X86::VMINPDrm_Int, TB_ALIGN_16 },
866 { X86::VMINPSrr, X86::VMINPSrm, TB_ALIGN_16 },
867 { X86::VMINPSrr_Int, X86::VMINPSrm_Int, TB_ALIGN_16 },
868 { X86::VMINSDrr, X86::VMINSDrm, 0 },
869 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 },
870 { X86::VMINSSrr, X86::VMINSSrm, 0 },
871 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]872 { X86::VMPSADBWrri, X86::VMPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]873 { X86::VMULPDrr, X86::VMULPDrm, TB_ALIGN_16 },
874 { X86::VMULPSrr, X86::VMULPSrm, TB_ALIGN_16 },
875 { X86::VMULSDrr, X86::VMULSDrm, 0 },
876 { X86::VMULSSrr, X86::VMULSSrm, 0 },
877 { X86::VORPDrr, X86::VORPDrm, TB_ALIGN_16 },
878 { X86::VORPSrr, X86::VORPSrm, TB_ALIGN_16 },
879 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, TB_ALIGN_16 },
880 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]881 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]882 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, TB_ALIGN_16 },
883 { X86::VPADDBrr, X86::VPADDBrm, TB_ALIGN_16 },
884 { X86::VPADDDrr, X86::VPADDDrm, TB_ALIGN_16 },
885 { X86::VPADDQrr, X86::VPADDQrm, TB_ALIGN_16 },
886 { X86::VPADDSBrr, X86::VPADDSBrm, TB_ALIGN_16 },
887 { X86::VPADDSWrr, X86::VPADDSWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]888 { X86::VPADDUSBrr, X86::VPADDUSBrm, TB_ALIGN_16 },
889 { X86::VPADDUSWrr, X86::VPADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]890 { X86::VPADDWrr, X86::VPADDWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]891 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]892 { X86::VPANDNrr, X86::VPANDNrm, TB_ALIGN_16 },
893 { X86::VPANDrr, X86::VPANDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]894 { X86::VPAVGBrr, X86::VPAVGBrm, TB_ALIGN_16 },
895 { X86::VPAVGWrr, X86::VPAVGWrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]896 { X86::VPBLENDWrri, X86::VPBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]897 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, TB_ALIGN_16 },
898 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]899 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]900 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, TB_ALIGN_16 },
901 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, TB_ALIGN_16 },
902 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]903 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]904 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]905 { X86::VPHADDDrr, X86::VPHADDDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]906 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]907 { X86::VPHADDWrr, X86::VPHADDWrm, TB_ALIGN_16 },
908 { X86::VPHSUBDrr, X86::VPHSUBDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]909 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]910 { X86::VPHSUBWrr, X86::VPHSUBWrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]911 { X86::VPERMILPDrr, X86::VPERMILPDrm, TB_ALIGN_16 },
912 { X86::VPERMILPSrr, X86::VPERMILPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]913 { X86::VPINSRWrri, X86::VPINSRWrmi, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]914 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]915 { X86::VPMADDWDrr, X86::VPMADDWDrm, TB_ALIGN_16 },
916 { X86::VPMAXSWrr, X86::VPMAXSWrm, TB_ALIGN_16 },
917 { X86::VPMAXUBrr, X86::VPMAXUBrm, TB_ALIGN_16 },
918 { X86::VPMINSWrr, X86::VPMINSWrm, TB_ALIGN_16 },
919 { X86::VPMINUBrr, X86::VPMINUBrm, TB_ALIGN_16 },
920 { X86::VPMULDQrr, X86::VPMULDQrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]921 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]922 { X86::VPMULHUWrr, X86::VPMULHUWrm, TB_ALIGN_16 },
923 { X86::VPMULHWrr, X86::VPMULHWrm, TB_ALIGN_16 },
924 { X86::VPMULLDrr, X86::VPMULLDrm, TB_ALIGN_16 },
925 { X86::VPMULLWrr, X86::VPMULLWrm, TB_ALIGN_16 },
926 { X86::VPMULUDQrr, X86::VPMULUDQrm, TB_ALIGN_16 },
927 { X86::VPORrr, X86::VPORrm, TB_ALIGN_16 },
928 { X86::VPSADBWrr, X86::VPSADBWrm, TB_ALIGN_16 },
Craig Topper969ba282012-01-25 06:43:11 +0000[diff] [blame]929 { X86::VPSHUFBrr, X86::VPSHUFBrm, TB_ALIGN_16 },
930 { X86::VPSIGNBrr, X86::VPSIGNBrm, TB_ALIGN_16 },
931 { X86::VPSIGNWrr, X86::VPSIGNWrm, TB_ALIGN_16 },
932 { X86::VPSIGNDrr, X86::VPSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]933 { X86::VPSLLDrr, X86::VPSLLDrm, TB_ALIGN_16 },
934 { X86::VPSLLQrr, X86::VPSLLQrm, TB_ALIGN_16 },
935 { X86::VPSLLWrr, X86::VPSLLWrm, TB_ALIGN_16 },
936 { X86::VPSRADrr, X86::VPSRADrm, TB_ALIGN_16 },
937 { X86::VPSRAWrr, X86::VPSRAWrm, TB_ALIGN_16 },
938 { X86::VPSRLDrr, X86::VPSRLDrm, TB_ALIGN_16 },
939 { X86::VPSRLQrr, X86::VPSRLQrm, TB_ALIGN_16 },
940 { X86::VPSRLWrr, X86::VPSRLWrm, TB_ALIGN_16 },
941 { X86::VPSUBBrr, X86::VPSUBBrm, TB_ALIGN_16 },
942 { X86::VPSUBDrr, X86::VPSUBDrm, TB_ALIGN_16 },
943 { X86::VPSUBSBrr, X86::VPSUBSBrm, TB_ALIGN_16 },
944 { X86::VPSUBSWrr, X86::VPSUBSWrm, TB_ALIGN_16 },
945 { X86::VPSUBWrr, X86::VPSUBWrm, TB_ALIGN_16 },
946 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, TB_ALIGN_16 },
947 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, TB_ALIGN_16 },
948 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, TB_ALIGN_16 },
949 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, TB_ALIGN_16 },
950 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, TB_ALIGN_16 },
951 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, TB_ALIGN_16 },
952 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, TB_ALIGN_16 },
953 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, TB_ALIGN_16 },
954 { X86::VPXORrr, X86::VPXORrm, TB_ALIGN_16 },
955 { X86::VSHUFPDrri, X86::VSHUFPDrmi, TB_ALIGN_16 },
956 { X86::VSHUFPSrri, X86::VSHUFPSrmi, TB_ALIGN_16 },
957 { X86::VSUBPDrr, X86::VSUBPDrm, TB_ALIGN_16 },
958 { X86::VSUBPSrr, X86::VSUBPSrm, TB_ALIGN_16 },
959 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
960 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
961 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, TB_ALIGN_16 },
962 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, TB_ALIGN_16 },
963 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, TB_ALIGN_16 },
964 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, TB_ALIGN_16 },
965 { X86::VXORPDrr, X86::VXORPDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]966 { X86::VXORPSrr, X86::VXORPSrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]967 // AVX 256-bit foldable instructions
968 { X86::VADDPDYrr, X86::VADDPDYrm, TB_ALIGN_32 },
969 { X86::VADDPSYrr, X86::VADDPSYrm, TB_ALIGN_32 },
970 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, TB_ALIGN_32 },
971 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, TB_ALIGN_32 },
972 { X86::VANDNPDYrr, X86::VANDNPDYrm, TB_ALIGN_32 },
973 { X86::VANDNPSYrr, X86::VANDNPSYrm, TB_ALIGN_32 },
974 { X86::VANDPDYrr, X86::VANDPDYrm, TB_ALIGN_32 },
975 { X86::VANDPSYrr, X86::VANDPSYrm, TB_ALIGN_32 },
976 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, TB_ALIGN_32 },
977 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, TB_ALIGN_32 },
978 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, TB_ALIGN_32 },
979 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, TB_ALIGN_32 },
980 { X86::VCMPPDYrri, X86::VCMPPDYrmi, TB_ALIGN_32 },
981 { X86::VCMPPSYrri, X86::VCMPPSYrmi, TB_ALIGN_32 },
982 { X86::VDIVPDYrr, X86::VDIVPDYrm, TB_ALIGN_32 },
983 { X86::VDIVPSYrr, X86::VDIVPSYrm, TB_ALIGN_32 },
984 { X86::VHADDPDYrr, X86::VHADDPDYrm, TB_ALIGN_32 },
985 { X86::VHADDPSYrr, X86::VHADDPSYrm, TB_ALIGN_32 },
986 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, TB_ALIGN_32 },
987 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, TB_ALIGN_32 },
988 { X86::VINSERTF128rr, X86::VINSERTF128rm, TB_ALIGN_32 },
989 { X86::VMAXPDYrr, X86::VMAXPDYrm, TB_ALIGN_32 },
990 { X86::VMAXPDYrr_Int, X86::VMAXPDYrm_Int, TB_ALIGN_32 },
991 { X86::VMAXPSYrr, X86::VMAXPSYrm, TB_ALIGN_32 },
992 { X86::VMAXPSYrr_Int, X86::VMAXPSYrm_Int, TB_ALIGN_32 },
993 { X86::VMINPDYrr, X86::VMINPDYrm, TB_ALIGN_32 },
994 { X86::VMINPDYrr_Int, X86::VMINPDYrm_Int, TB_ALIGN_32 },
995 { X86::VMINPSYrr, X86::VMINPSYrm, TB_ALIGN_32 },
996 { X86::VMINPSYrr_Int, X86::VMINPSYrm_Int, TB_ALIGN_32 },
997 { X86::VMULPDYrr, X86::VMULPDYrm, TB_ALIGN_32 },
998 { X86::VMULPSYrr, X86::VMULPSYrm, TB_ALIGN_32 },
999 { X86::VORPDYrr, X86::VORPDYrm, TB_ALIGN_32 },
1000 { X86::VORPSYrr, X86::VORPSYrm, TB_ALIGN_32 },
1001 { X86::VPERM2F128rr, X86::VPERM2F128rm, TB_ALIGN_32 },
1002 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, TB_ALIGN_32 },
1003 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, TB_ALIGN_32 },
1004 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, TB_ALIGN_32 },
1005 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, TB_ALIGN_32 },
1006 { X86::VSUBPDYrr, X86::VSUBPDYrm, TB_ALIGN_32 },
1007 { X86::VSUBPSYrr, X86::VSUBPSYrm, TB_ALIGN_32 },
1008 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, TB_ALIGN_32 },
1009 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, TB_ALIGN_32 },
1010 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, TB_ALIGN_32 },
1011 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, TB_ALIGN_32 },
1012 { X86::VXORPDYrr, X86::VXORPDYrm, TB_ALIGN_32 },
1013 { X86::VXORPSYrr, X86::VXORPSYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1014 // AVX2 foldable instructions
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1015 { X86::VINSERTI128rr, X86::VINSERTI128rm, TB_ALIGN_16 },
1016 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, TB_ALIGN_32 },
1017 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, TB_ALIGN_32 },
1018 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, TB_ALIGN_32 },
1019 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, TB_ALIGN_32 },
1020 { X86::VPADDBYrr, X86::VPADDBYrm, TB_ALIGN_32 },
1021 { X86::VPADDDYrr, X86::VPADDDYrm, TB_ALIGN_32 },
1022 { X86::VPADDQYrr, X86::VPADDQYrm, TB_ALIGN_32 },
1023 { X86::VPADDSBYrr, X86::VPADDSBYrm, TB_ALIGN_32 },
1024 { X86::VPADDSWYrr, X86::VPADDSWYrm, TB_ALIGN_32 },
1025 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, TB_ALIGN_32 },
1026 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, TB_ALIGN_32 },
1027 { X86::VPADDWYrr, X86::VPADDWYrm, TB_ALIGN_32 },
1028 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, TB_ALIGN_32 },
1029 { X86::VPANDNYrr, X86::VPANDNYrm, TB_ALIGN_32 },
1030 { X86::VPANDYrr, X86::VPANDYrm, TB_ALIGN_32 },
1031 { X86::VPAVGBYrr, X86::VPAVGBYrm, TB_ALIGN_32 },
1032 { X86::VPAVGWYrr, X86::VPAVGWYrm, TB_ALIGN_32 },
1033 { X86::VPBLENDDrri, X86::VPBLENDDrmi, TB_ALIGN_32 },
1034 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, TB_ALIGN_32 },
1035 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, TB_ALIGN_32 },
1036 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, TB_ALIGN_32 },
1037 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, TB_ALIGN_32 },
1038 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, TB_ALIGN_32 },
1039 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, TB_ALIGN_32 },
1040 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, TB_ALIGN_32 },
1041 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, TB_ALIGN_32 },
1042 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, TB_ALIGN_32 },
1043 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, TB_ALIGN_32 },
1044 { X86::VPERM2I128rr, X86::VPERM2I128rm, TB_ALIGN_32 },
Craig Topper40385c82012-01-19 08:50:38 +0000[diff] [blame]1045 { X86::VPERMDYrr, X86::VPERMDYrm, TB_ALIGN_32 },
Elena Demikhovsky73c504a2012-04-15 11:18:59 +0000[diff] [blame]1046 { X86::VPERMPDYri, X86::VPERMPDYmi, TB_ALIGN_32 },
Craig Topper40385c82012-01-19 08:50:38 +0000[diff] [blame]1047 { X86::VPERMPSYrr, X86::VPERMPSYrm, TB_ALIGN_32 },
Elena Demikhovsky73c504a2012-04-15 11:18:59 +0000[diff] [blame]1048 { X86::VPERMQYri, X86::VPERMQYmi, TB_ALIGN_32 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]1049 { X86::VPHADDDYrr, X86::VPHADDDYrm, TB_ALIGN_32 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1050 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, TB_ALIGN_32 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]1051 { X86::VPHADDWYrr, X86::VPHADDWYrm, TB_ALIGN_32 },
1052 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, TB_ALIGN_32 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1053 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, TB_ALIGN_32 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]1054 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, TB_ALIGN_32 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1055 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, TB_ALIGN_32 },
1056 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, TB_ALIGN_32 },
1057 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, TB_ALIGN_32 },
1058 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, TB_ALIGN_32 },
1059 { X86::VPMINSWYrr, X86::VPMINSWYrm, TB_ALIGN_32 },
1060 { X86::VPMINUBYrr, X86::VPMINUBYrm, TB_ALIGN_32 },
1061 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, TB_ALIGN_32 },
1062 { X86::VPMULDQYrr, X86::VPMULDQYrm, TB_ALIGN_32 },
1063 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, TB_ALIGN_32 },
1064 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, TB_ALIGN_32 },
1065 { X86::VPMULHWYrr, X86::VPMULHWYrm, TB_ALIGN_32 },
1066 { X86::VPMULLDYrr, X86::VPMULLDYrm, TB_ALIGN_32 },
1067 { X86::VPMULLWYrr, X86::VPMULLWYrm, TB_ALIGN_32 },
1068 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, TB_ALIGN_32 },
1069 { X86::VPORYrr, X86::VPORYrm, TB_ALIGN_32 },
1070 { X86::VPSADBWYrr, X86::VPSADBWYrm, TB_ALIGN_32 },
Craig Topper969ba282012-01-25 06:43:11 +0000[diff] [blame]1071 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, TB_ALIGN_32 },
1072 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, TB_ALIGN_32 },
1073 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, TB_ALIGN_32 },
1074 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1075 { X86::VPSLLDYrr, X86::VPSLLDYrm, TB_ALIGN_16 },
1076 { X86::VPSLLQYrr, X86::VPSLLQYrm, TB_ALIGN_16 },
1077 { X86::VPSLLWYrr, X86::VPSLLWYrm, TB_ALIGN_16 },
1078 { X86::VPSLLVDrr, X86::VPSLLVDrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1079 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1080 { X86::VPSLLVQrr, X86::VPSLLVQrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1081 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1082 { X86::VPSRADYrr, X86::VPSRADYrm, TB_ALIGN_16 },
1083 { X86::VPSRAWYrr, X86::VPSRAWYrm, TB_ALIGN_16 },
1084 { X86::VPSRAVDrr, X86::VPSRAVDrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1085 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1086 { X86::VPSRLDYrr, X86::VPSRLDYrm, TB_ALIGN_16 },
1087 { X86::VPSRLQYrr, X86::VPSRLQYrm, TB_ALIGN_16 },
1088 { X86::VPSRLWYrr, X86::VPSRLWYrm, TB_ALIGN_16 },
1089 { X86::VPSRLVDrr, X86::VPSRLVDrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1090 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1091 { X86::VPSRLVQrr, X86::VPSRLVQrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1092 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, TB_ALIGN_32 },
1093 { X86::VPSUBBYrr, X86::VPSUBBYrm, TB_ALIGN_32 },
1094 { X86::VPSUBDYrr, X86::VPSUBDYrm, TB_ALIGN_32 },
1095 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, TB_ALIGN_32 },
1096 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, TB_ALIGN_32 },
1097 { X86::VPSUBWYrr, X86::VPSUBWYrm, TB_ALIGN_32 },
1098 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, TB_ALIGN_32 },
1099 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1100 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1101 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, TB_ALIGN_32 },
1102 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, TB_ALIGN_32 },
1103 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, TB_ALIGN_32 },
1104 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, TB_ALIGN_32 },
1105 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, TB_ALIGN_32 },
1106 { X86::VPXORYrr, X86::VPXORYrm, TB_ALIGN_32 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]1107 // FIXME: add AVX 256-bit foldable instructions
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]1108 };
1109
1110 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]1111 unsigned RegOp = OpTbl2[i].RegOp;
1112 unsigned MemOp = OpTbl2[i].MemOp;
1113 unsigned Flags = OpTbl2[i].Flags;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]1114 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1115 RegOp, MemOp,
1116 // Index 2, folded load
1117 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]1118 }
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1119
1120 static const X86OpTblEntry OpTbl3[] = {
1121 // FMA foldable instructions
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1122 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1123 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1124 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1125 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1126 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1127 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
1128 { X86::VFMADDSSr132r_Int, X86::VFMADDSSr132m_Int, 0 },
1129 { X86::VFMADDSDr132r_Int, X86::VFMADDSDr132m_Int, 0 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1130
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1131 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1132 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1133 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1134 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1135 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1136 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1137 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1138 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1139 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1140 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1141 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1142 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
1143 { X86::VFMADDPSr132r_Int, X86::VFMADDPSr132m_Int, TB_ALIGN_16 },
1144 { X86::VFMADDPDr132r_Int, X86::VFMADDPDr132m_Int, TB_ALIGN_16 },
1145 { X86::VFMADDPSr132rY_Int, X86::VFMADDPSr132mY_Int, TB_ALIGN_32 },
1146 { X86::VFMADDPDr132rY_Int, X86::VFMADDPDr132mY_Int, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1147
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1148 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1149 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1150 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1151 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1152 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1153 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
1154 { X86::VFNMADDSSr132r_Int, X86::VFNMADDSSr132m_Int, 0 },
1155 { X86::VFNMADDSDr132r_Int, X86::VFNMADDSDr132m_Int, 0 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1156
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1157 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1158 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1159 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1160 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1161 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1162 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1163 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1164 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1165 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1166 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1167 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1168 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
1169 { X86::VFNMADDPSr132r_Int, X86::VFNMADDPSr132m_Int, TB_ALIGN_16 },
1170 { X86::VFNMADDPDr132r_Int, X86::VFNMADDPDr132m_Int, TB_ALIGN_16 },
1171 { X86::VFNMADDPSr132rY_Int, X86::VFNMADDPSr132mY_Int, TB_ALIGN_32 },
1172 { X86::VFNMADDPDr132rY_Int, X86::VFNMADDPDr132mY_Int, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1173
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1174 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1175 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1176 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1177 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1178 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1179 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
1180 { X86::VFMSUBSSr132r_Int, X86::VFMSUBSSr132m_Int, 0 },
1181 { X86::VFMSUBSDr132r_Int, X86::VFMSUBSDr132m_Int, 0 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1182
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1183 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1184 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1185 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1186 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1187 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1188 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1189 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1190 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1191 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1192 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1193 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1194 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
1195 { X86::VFMSUBPSr132r_Int, X86::VFMSUBPSr132m_Int, TB_ALIGN_16 },
1196 { X86::VFMSUBPDr132r_Int, X86::VFMSUBPDr132m_Int, TB_ALIGN_16 },
1197 { X86::VFMSUBPSr132rY_Int, X86::VFMSUBPSr132mY_Int, TB_ALIGN_32 },
1198 { X86::VFMSUBPDr132rY_Int, X86::VFMSUBPDr132mY_Int, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1199
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1200 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1201 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1202 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1203 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1204 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1205 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
1206 { X86::VFNMSUBSSr132r_Int, X86::VFNMSUBSSr132m_Int, 0 },
1207 { X86::VFNMSUBSDr132r_Int, X86::VFNMSUBSDr132m_Int, 0 },
Craig Topper78fc72d2012-06-01 05:48:39 +0000[diff] [blame]1208
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1209 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1210 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1211 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1212 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1213 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1214 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1215 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1216 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1217 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1218 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1219 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1220 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
1221 { X86::VFNMSUBPSr132r_Int, X86::VFNMSUBPSr132m_Int, TB_ALIGN_16 },
1222 { X86::VFNMSUBPDr132r_Int, X86::VFNMSUBPDr132m_Int, TB_ALIGN_16 },
1223 { X86::VFNMSUBPSr132rY_Int, X86::VFNMSUBPSr132mY_Int, TB_ALIGN_32 },
1224 { X86::VFNMSUBPDr132rY_Int, X86::VFNMSUBPDr132mY_Int, TB_ALIGN_32 },
Craig Topperfc5ab242012-06-04 07:08:21 +0000[diff] [blame]1225
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1226 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1227 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1228 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1229 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1230 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1231 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1232 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1233 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1234 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1235 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1236 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1237 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
1238 { X86::VFMADDSUBPSr132r_Int, X86::VFMADDSUBPSr132m_Int, TB_ALIGN_16 },
1239 { X86::VFMADDSUBPDr132r_Int, X86::VFMADDSUBPDr132m_Int, TB_ALIGN_16 },
1240 { X86::VFMADDSUBPSr132rY_Int, X86::VFMADDSUBPSr132mY_Int, TB_ALIGN_32 },
1241 { X86::VFMADDSUBPDr132rY_Int, X86::VFMADDSUBPDr132mY_Int, TB_ALIGN_32 },
Craig Topperfc5ab242012-06-04 07:08:21 +0000[diff] [blame]1242
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1243 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1244 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1245 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1246 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1247 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1248 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1249 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1250 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1251 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1252 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1253 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1254 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
1255 { X86::VFMSUBADDPSr132r_Int, X86::VFMSUBADDPSr132m_Int, TB_ALIGN_16 },
1256 { X86::VFMSUBADDPDr132r_Int, X86::VFMSUBADDPDr132m_Int, TB_ALIGN_16 },
1257 { X86::VFMSUBADDPSr132rY_Int, X86::VFMSUBADDPSr132mY_Int, TB_ALIGN_32 },
1258 { X86::VFMSUBADDPDr132rY_Int, X86::VFMSUBADDPDr132mY_Int, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1259 };
1260
1261 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1262 unsigned RegOp = OpTbl3[i].RegOp;
1263 unsigned MemOp = OpTbl3[i].MemOp;
1264 unsigned Flags = OpTbl3[i].Flags;
1265 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1266 RegOp, MemOp,
1267 // Index 3, folded load
1268 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1269 }
1270
Chris Lattner72614082002-10-25 22:55:53 +0000[diff] [blame]1271}
1272
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]1273void
1274X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1275 MemOp2RegOpTableType &M2RTable,
1276 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1277 if ((Flags & TB_NO_FORWARD) == 0) {
1278 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1279 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1280 }
1281 if ((Flags & TB_NO_REVERSE) == 0) {
1282 assert(!M2RTable.count(MemOp) &&
1283 "Duplicated entries in unfolding maps?");
1284 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1285 }
1286}
1287
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1288bool
Evan Cheng7da9ecf2010-01-13 00:30:23 +0000[diff] [blame]1289X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1290 unsigned &SrcReg, unsigned &DstReg,
1291 unsigned &SubIdx) const {
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1292 switch (MI.getOpcode()) {
1293 default: break;
1294 case X86::MOVSX16rr8:
1295 case X86::MOVZX16rr8:
1296 case X86::MOVSX32rr8:
1297 case X86::MOVZX32rr8:
1298 case X86::MOVSX64rr8:
1299 case X86::MOVZX64rr8:
Evan Cheng57d1d932010-01-13 08:01:32 +0000[diff] [blame]1300 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1301 // It's not always legal to reference the low 8-bit of the larger
1302 // register in 32-bit mode.
1303 return false;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1304 case X86::MOVSX32rr16:
1305 case X86::MOVZX32rr16:
1306 case X86::MOVSX64rr16:
1307 case X86::MOVZX64rr16:
1308 case X86::MOVSX64rr32:
1309 case X86::MOVZX64rr32: {
1310 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1311 // Be conservative.
1312 return false;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1313 SrcReg = MI.getOperand(1).getReg();
1314 DstReg = MI.getOperand(0).getReg();
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1315 switch (MI.getOpcode()) {
1316 default:
1317 llvm_unreachable(0);
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1318 case X86::MOVSX16rr8:
1319 case X86::MOVZX16rr8:
1320 case X86::MOVSX32rr8:
1321 case X86::MOVZX32rr8:
1322 case X86::MOVSX64rr8:
1323 case X86::MOVZX64rr8:
Jakob Stoklund Olesen22c0e972010-05-25 17:04:16 +0000[diff] [blame]1324 SubIdx = X86::sub_8bit;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1325 break;
1326 case X86::MOVSX32rr16:
1327 case X86::MOVZX32rr16:
1328 case X86::MOVSX64rr16:
1329 case X86::MOVZX64rr16:
Jakob Stoklund Olesen22c0e972010-05-25 17:04:16 +0000[diff] [blame]1330 SubIdx = X86::sub_16bit;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1331 break;
1332 case X86::MOVSX64rr32:
1333 case X86::MOVZX64rr32:
Jakob Stoklund Olesen22c0e972010-05-25 17:04:16 +0000[diff] [blame]1334 SubIdx = X86::sub_32bit;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1335 break;
1336 }
Evan Cheng7da9ecf2010-01-13 00:30:23 +0000[diff] [blame]1337 return true;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1338 }
1339 }
Evan Cheng7da9ecf2010-01-13 00:30:23 +0000[diff] [blame]1340 return false;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1341}
1342
David Greeneb87bc952009-11-12 20:55:29 +0000[diff] [blame]1343/// isFrameOperand - Return true and the FrameIndex if the specified
1344/// operand and follow operands form a reference to the stack frame.
1345bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1346 int &FrameIndex) const {
1347 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1348 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1349 MI->getOperand(Op+1).getImm() == 1 &&
1350 MI->getOperand(Op+2).getReg() == 0 &&
1351 MI->getOperand(Op+3).getImm() == 0) {
1352 FrameIndex = MI->getOperand(Op).getIndex();
1353 return true;
1354 }
1355 return false;
1356}
1357
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1358static bool isFrameLoadOpcode(int Opcode) {
1359 switch (Opcode) {
David Blaikie4d6ccb52012-01-20 21:51:11 +0000[diff] [blame]1360 default:
1361 return false;
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1362 case X86::MOV8rm:
1363 case X86::MOV16rm:
1364 case X86::MOV32rm:
Evan Cheng25ab6902006-09-08 06:48:29 +0000[diff] [blame]1365 case X86::MOV64rm:
Dale Johannesene377d4d2007-07-04 21:07:47 +0000[diff] [blame]1366 case X86::LD_Fp64m:
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1367 case X86::MOVSSrm:
1368 case X86::MOVSDrm:
Chris Lattner993c8972006-04-18 16:44:51 +0000[diff] [blame]1369 case X86::MOVAPSrm:
1370 case X86::MOVAPDrm:
Dan Gohman54462742009-01-09 02:40:34 +0000[diff] [blame]1371 case X86::MOVDQArm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]1372 case X86::VMOVSSrm:
1373 case X86::VMOVSDrm:
1374 case X86::VMOVAPSrm:
1375 case X86::VMOVAPDrm:
1376 case X86::VMOVDQArm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]1377 case X86::VMOVAPSYrm:
1378 case X86::VMOVAPDYrm:
1379 case X86::VMOVDQAYrm:
Bill Wendling823efee2007-04-03 06:00:37 +0000[diff] [blame]1380 case X86::MMX_MOVD64rm:
1381 case X86::MMX_MOVQ64rm:
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1382 return true;
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1383 }
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1384}
1385
1386static bool isFrameStoreOpcode(int Opcode) {
1387 switch (Opcode) {
1388 default: break;
1389 case X86::MOV8mr:
1390 case X86::MOV16mr:
1391 case X86::MOV32mr:
1392 case X86::MOV64mr:
1393 case X86::ST_FpP64m:
1394 case X86::MOVSSmr:
1395 case X86::MOVSDmr:
1396 case X86::MOVAPSmr:
1397 case X86::MOVAPDmr:
1398 case X86::MOVDQAmr:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]1399 case X86::VMOVSSmr:
1400 case X86::VMOVSDmr:
1401 case X86::VMOVAPSmr:
1402 case X86::VMOVAPDmr:
1403 case X86::VMOVDQAmr:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]1404 case X86::VMOVAPSYmr:
1405 case X86::VMOVAPDYmr:
1406 case X86::VMOVDQAYmr:
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1407 case X86::MMX_MOVD64mr:
1408 case X86::MMX_MOVQ64mr:
1409 case X86::MMX_MOVNTQmr:
1410 return true;
1411 }
1412 return false;
1413}
1414
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1415unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1416 int &FrameIndex) const {
1417 if (isFrameLoadOpcode(MI->getOpcode()))
Jakob Stoklund Olesen81c7b192010-07-27 04:17:01 +0000[diff] [blame]1418 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1419 return MI->getOperand(0).getReg();
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1420 return 0;
1421}
1422
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1423unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1424 int &FrameIndex) const {
1425 if (isFrameLoadOpcode(MI->getOpcode())) {
1426 unsigned Reg;
1427 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1428 return Reg;
David Greeneb87bc952009-11-12 20:55:29 +0000[diff] [blame]1429 // Check for post-frame index elimination operations
David Greene29dbf502009-12-04 22:38:46 +0000[diff] [blame]1430 const MachineMemOperand *Dummy;
1431 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1432 }
1433 return 0;
1434}
1435
Dan Gohmancbad42c2008-11-18 19:49:32 +0000[diff] [blame]1436unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1437 int &FrameIndex) const {
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1438 if (isFrameStoreOpcode(MI->getOpcode()))
Jakob Stoklund Olesen81c7b192010-07-27 04:17:01 +0000[diff] [blame]1439 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1440 isFrameOperand(MI, 0, FrameIndex))
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]1441 return MI->getOperand(X86::AddrNumOperands).getReg();
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1442 return 0;
1443}
1444
1445unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1446 int &FrameIndex) const {
1447 if (isFrameStoreOpcode(MI->getOpcode())) {
1448 unsigned Reg;
1449 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1450 return Reg;
David Greeneb87bc952009-11-12 20:55:29 +0000[diff] [blame]1451 // Check for post-frame index elimination operations
David Greene29dbf502009-12-04 22:38:46 +0000[diff] [blame]1452 const MachineMemOperand *Dummy;
1453 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1454 }
1455 return 0;
1456}
1457
Evan Chenge3d8dbf2008-03-27 01:45:11 +0000[diff] [blame]1458/// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1459/// X86::MOVPC32r.
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]1460static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
Evan Chenge3d8dbf2008-03-27 01:45:11 +0000[diff] [blame]1461 bool isPICBase = false;
1462 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1463 E = MRI.def_end(); I != E; ++I) {
1464 MachineInstr *DefMI = I.getOperand().getParent();
1465 if (DefMI->getOpcode() != X86::MOVPC32r)
1466 return false;
1467 assert(!isPICBase && "More than one PIC base?");
1468 isPICBase = true;
1469 }
1470 return isPICBase;
1471}
Evan Cheng9d15abe2008-03-31 07:54:19 +0000[diff] [blame]1472
Bill Wendling9f8fea32008-05-12 20:54:26 +0000[diff] [blame]1473bool
Dan Gohman3731bc02009-10-10 00:34:18 +0000[diff] [blame]1474X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1475 AliasAnalysis *AA) const {
Dan Gohmanc101e952007-06-14 20:50:44 +0000[diff] [blame]1476 switch (MI->getOpcode()) {
1477 default: break;
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1478 case X86::MOV8rm:
1479 case X86::MOV16rm:
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1480 case X86::MOV32rm:
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1481 case X86::MOV64rm:
1482 case X86::LD_Fp64m:
1483 case X86::MOVSSrm:
1484 case X86::MOVSDrm:
1485 case X86::MOVAPSrm:
Evan Cheng600c0432009-11-16 21:56:03 +0000[diff] [blame]1486 case X86::MOVUPSrm:
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1487 case X86::MOVAPDrm:
Dan Gohman54462742009-01-09 02:40:34 +0000[diff] [blame]1488 case X86::MOVDQArm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]1489 case X86::VMOVSSrm:
1490 case X86::VMOVSDrm:
1491 case X86::VMOVAPSrm:
1492 case X86::VMOVUPSrm:
1493 case X86::VMOVAPDrm:
1494 case X86::VMOVDQArm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]1495 case X86::VMOVAPSYrm:
1496 case X86::VMOVUPSYrm:
1497 case X86::VMOVAPDYrm:
1498 case X86::VMOVDQAYrm:
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1499 case X86::MMX_MOVD64rm:
Evan Chengd15ac2f2009-11-17 09:51:18 +0000[diff] [blame]1500 case X86::MMX_MOVQ64rm:
Bruno Cardoso Lopes0e59a042011-09-03 00:46:45 +0000[diff] [blame]1501 case X86::FsVMOVAPSrm:
1502 case X86::FsVMOVAPDrm:
Evan Chengd15ac2f2009-11-17 09:51:18 +0000[diff] [blame]1503 case X86::FsMOVAPSrm:
1504 case X86::FsMOVAPDrm: {
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1505 // Loads from constant pools are trivially rematerializable.
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]1506 if (MI->getOperand(1).isReg() &&
1507 MI->getOperand(2).isImm() &&
1508 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
Dan Gohman3731bc02009-10-10 00:34:18 +0000[diff] [blame]1509 MI->isInvariantLoad(AA)) {
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1510 unsigned BaseReg = MI->getOperand(1).getReg();
Chris Lattner18c59872009-06-27 04:16:01 +0000[diff] [blame]1511 if (BaseReg == 0 || BaseReg == X86::RIP)
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1512 return true;
1513 // Allow re-materialization of PIC load.
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]1514 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
Evan Chengffe2eb02008-04-01 23:26:12 +0000[diff] [blame]1515 return false;
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]1516 const MachineFunction &MF = *MI->getParent()->getParent();
1517 const MachineRegisterInfo &MRI = MF.getRegInfo();
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1518 bool isPICBase = false;
1519 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1520 E = MRI.def_end(); I != E; ++I) {
1521 MachineInstr *DefMI = I.getOperand().getParent();
1522 if (DefMI->getOpcode() != X86::MOVPC32r)
1523 return false;
1524 assert(!isPICBase && "More than one PIC base?");
1525 isPICBase = true;
1526 }
1527 return isPICBase;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1528 }
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1529 return false;
Evan Chengd8850a52008-02-22 09:25:47 +0000[diff] [blame]1530 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1531
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1532 case X86::LEA32r:
1533 case X86::LEA64r: {
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]1534 if (MI->getOperand(2).isImm() &&
1535 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1536 !MI->getOperand(4).isReg()) {
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1537 // lea fi#, lea GV, etc. are all rematerializable.
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]1538 if (!MI->getOperand(1).isReg())
Dan Gohman83ccd142008-09-26 21:30:20 +0000[diff] [blame]1539 return true;
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1540 unsigned BaseReg = MI->getOperand(1).getReg();
1541 if (BaseReg == 0)
1542 return true;
1543 // Allow re-materialization of lea PICBase + x.
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]1544 const MachineFunction &MF = *MI->getParent()->getParent();
1545 const MachineRegisterInfo &MRI = MF.getRegInfo();
Evan Chenge3d8dbf2008-03-27 01:45:11 +0000[diff] [blame]1546 return regIsPICBase(BaseReg, MRI);
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1547 }
1548 return false;
1549 }
Dan Gohmanc101e952007-06-14 20:50:44 +0000[diff] [blame]1550 }
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1551
Dan Gohmand45eddd2007-06-26 00:48:07 +0000[diff] [blame]1552 // All other instructions marked M_REMATERIALIZABLE are always trivially
1553 // rematerializable.
1554 return true;
Dan Gohmanc101e952007-06-14 20:50:44 +0000[diff] [blame]1555}
1556
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1557/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1558/// would clobber the EFLAGS condition register. Note the result may be
1559/// conservative. If it cannot definitely determine the safety after visiting
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1560/// a few instructions in each direction it assumes it's not safe.
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1561static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1562 MachineBasicBlock::iterator I) {
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1563 MachineBasicBlock::iterator E = MBB.end();
1564
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1565 // For compile time consideration, if we are not able to determine the
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1566 // safety after visiting 4 instructions in each direction, we will assume
1567 // it's not safe.
1568 MachineBasicBlock::iterator Iter = I;
Jakob Stoklund Olesenb8e052e2011-09-02 23:52:52 +0000[diff] [blame]1569 for (unsigned i = 0; Iter != E && i < 4; ++i) {
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1570 bool SeenDef = false;
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1571 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1572 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen450b3852012-02-09 00:17:22 +0000[diff] [blame]1573 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1574 SeenDef = true;
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]1575 if (!MO.isReg())
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1576 continue;
1577 if (MO.getReg() == X86::EFLAGS) {
1578 if (MO.isUse())
1579 return false;
1580 SeenDef = true;
1581 }
1582 }
1583
1584 if (SeenDef)
1585 // This instruction defines EFLAGS, no need to look any further.
1586 return true;
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1587 ++Iter;
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1588 // Skip over DBG_VALUE.
1589 while (Iter != E && Iter->isDebugValue())
1590 ++Iter;
Jakob Stoklund Olesenb8e052e2011-09-02 23:52:52 +0000[diff] [blame]1591 }
Dan Gohman3afda6e2008-10-21 03:24:31 +0000[diff] [blame]1592
Jakob Stoklund Olesenb8e052e2011-09-02 23:52:52 +0000[diff] [blame]1593 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1594 // live in.
1595 if (Iter == E) {
1596 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1597 SE = MBB.succ_end(); SI != SE; ++SI)
1598 if ((*SI)->isLiveIn(X86::EFLAGS))
1599 return false;
1600 return true;
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1601 }
1602
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1603 MachineBasicBlock::iterator B = MBB.begin();
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1604 Iter = I;
1605 for (unsigned i = 0; i < 4; ++i) {
1606 // If we make it to the beginning of the block, it's safe to clobber
1607 // EFLAGS iff EFLAGS is not live-in.
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1608 if (Iter == B)
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1609 return !MBB.isLiveIn(X86::EFLAGS);
1610
1611 --Iter;
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1612 // Skip over DBG_VALUE.
1613 while (Iter != B && Iter->isDebugValue())
1614 --Iter;
1615
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1616 bool SawKill = false;
1617 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1618 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen450b3852012-02-09 00:17:22 +0000[diff] [blame]1619 // A register mask may clobber EFLAGS, but we should still look for a
1620 // live EFLAGS def.
1621 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1622 SawKill = true;
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1623 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1624 if (MO.isDef()) return MO.isDead();
1625 if (MO.isKill()) SawKill = true;
1626 }
1627 }
1628
1629 if (SawKill)
1630 // This instruction kills EFLAGS and doesn't redefine it, so
1631 // there's no need to look further.
Dan Gohman3afda6e2008-10-21 03:24:31 +0000[diff] [blame]1632 return true;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1633 }
1634
1635 // Conservative answer.
1636 return false;
1637}
1638
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1639void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1640 MachineBasicBlock::iterator I,
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1641 unsigned DestReg, unsigned SubIdx,
Evan Chengd57cdd52009-11-14 02:55:43 +0000[diff] [blame]1642 const MachineInstr *Orig,
Jakob Stoklund Olesen9edf7de2010-06-02 22:47:25 +0000[diff] [blame]1643 const TargetRegisterInfo &TRI) const {
Dan Gohman0d881042010-05-07 01:28:10 +0000[diff] [blame]1644 DebugLoc DL = Orig->getDebugLoc();
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1645
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1646 // MOV32r0 etc. are implemented with xor which clobbers condition code.
1647 // Re-materialize them as movri instructions to avoid side effects.
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1648 bool Clone = true;
1649 unsigned Opc = Orig->getOpcode();
1650 switch (Opc) {
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1651 default: break;
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1652 case X86::MOV8r0:
Dan Gohmanf1b4d262010-01-12 04:42:54 +0000[diff] [blame]1653 case X86::MOV16r0:
1654 case X86::MOV32r0:
1655 case X86::MOV64r0: {
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1656 if (!isSafeToClobberEFLAGS(MBB, I)) {
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1657 switch (Opc) {
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1658 default: break;
1659 case X86::MOV8r0: Opc = X86::MOV8ri; break;
Dan Gohmanf1b4d262010-01-12 04:42:54 +0000[diff] [blame]1660 case X86::MOV16r0: Opc = X86::MOV16ri; break;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1661 case X86::MOV32r0: Opc = X86::MOV32ri; break;
Dan Gohman6fe0df22010-02-26 16:49:27 +0000[diff] [blame]1662 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1663 }
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1664 Clone = false;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1665 }
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1666 break;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1667 }
1668 }
1669
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1670 if (Clone) {
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]1671 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1672 MBB.insert(I, MI);
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1673 } else {
Jakob Stoklund Olesen9edf7de2010-06-02 22:47:25 +0000[diff] [blame]1674 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0);
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1675 }
Evan Cheng03eb3882008-04-16 23:44:44 +0000[diff] [blame]1676
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1677 MachineInstr *NewMI = prior(I);
Jakob Stoklund Olesen9edf7de2010-06-02 22:47:25 +0000[diff] [blame]1678 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1679}
1680
Evan Cheng3f411c72007-10-05 08:04:01 +0000[diff] [blame]1681/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1682/// is not marked dead.
1683static bool hasLiveCondCodeDef(MachineInstr *MI) {
Evan Cheng3f411c72007-10-05 08:04:01 +0000[diff] [blame]1684 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1685 MachineOperand &MO = MI->getOperand(i);
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]1686 if (MO.isReg() && MO.isDef() &&
Evan Cheng3f411c72007-10-05 08:04:01 +0000[diff] [blame]1687 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1688 return true;
1689 }
1690 }
1691 return false;
1692}
1693
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1694/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1695/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1696/// to a 32-bit superregister and then truncating back down to a 16-bit
1697/// subregister.
1698MachineInstr *
1699X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1700 MachineFunction::iterator &MFI,
1701 MachineBasicBlock::iterator &MBBI,
1702 LiveVariables *LV) const {
1703 MachineInstr *MI = MBBI;
1704 unsigned Dest = MI->getOperand(0).getReg();
1705 unsigned Src = MI->getOperand(1).getReg();
1706 bool isDead = MI->getOperand(0).isDead();
1707 bool isKill = MI->getOperand(1).isKill();
1708
1709 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1710 ? X86::LEA64_32r : X86::LEA32r;
1711 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1712 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1713 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1714
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1715 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1716 // well be shifting and then extracting the lower 16-bits.
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1717 // This has the potential to cause partial register stall. e.g.
Evan Cheng04ab19c2009-12-12 18:55:26 +0000[diff] [blame]1718 // movw (%rbp,%rcx,2), %dx
1719 // leal -65(%rdx), %esi
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1720 // But testing has shown this *does* help performance in 64-bit mode (at
1721 // least on modern x86 machines).
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1722 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1723 MachineInstr *InsMI =
Jakob Stoklund Olesen5c00e072010-07-08 16:40:15 +0000[diff] [blame]1724 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1725 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1726 .addReg(Src, getKillRegState(isKill));
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1727
1728 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1729 get(Opc), leaOutReg);
1730 switch (MIOpc) {
1731 default:
1732 llvm_unreachable(0);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1733 case X86::SHL16ri: {
1734 unsigned ShAmt = MI->getOperand(2).getImm();
1735 MIB.addReg(0).addImm(1 << ShAmt)
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1736 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1737 break;
1738 }
1739 case X86::INC16r:
1740 case X86::INC64_16r:
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1741 addRegOffset(MIB, leaInReg, true, 1);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1742 break;
1743 case X86::DEC16r:
1744 case X86::DEC64_16r:
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1745 addRegOffset(MIB, leaInReg, true, -1);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1746 break;
1747 case X86::ADD16ri:
1748 case X86::ADD16ri8:
Chris Lattner15df55d2010-10-08 03:57:25 +0000[diff] [blame]1749 case X86::ADD16ri_DB:
1750 case X86::ADD16ri8_DB:
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1751 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1752 break;
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]1753 case X86::ADD16rr:
1754 case X86::ADD16rr_DB: {
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1755 unsigned Src2 = MI->getOperand(2).getReg();
1756 bool isKill2 = MI->getOperand(2).isKill();
1757 unsigned leaInReg2 = 0;
1758 MachineInstr *InsMI2 = 0;
1759 if (Src == Src2) {
1760 // ADD16rr %reg1028<kill>, %reg1028
1761 // just a single insert_subreg.
1762 addRegReg(MIB, leaInReg, true, leaInReg, false);
1763 } else {
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1764 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1765 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1766 // well be shifting and then extracting the lower 16-bits.
Evan Chengddfd1372011-12-14 02:11:42 +0000[diff] [blame]1767 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1768 InsMI2 =
Evan Chengddfd1372011-12-14 02:11:42 +0000[diff] [blame]1769 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
Jakob Stoklund Olesen5c00e072010-07-08 16:40:15 +0000[diff] [blame]1770 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1771 .addReg(Src2, getKillRegState(isKill2));
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1772 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1773 }
1774 if (LV && isKill2 && InsMI2)
1775 LV->replaceKillInstruction(Src2, MI, InsMI2);
1776 break;
1777 }
1778 }
1779
1780 MachineInstr *NewMI = MIB;
1781 MachineInstr *ExtMI =
Jakob Stoklund Olesen0bc25f42010-07-08 16:40:22 +0000[diff] [blame]1782 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1783 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Jakob Stoklund Olesen0bc25f42010-07-08 16:40:22 +0000[diff] [blame]1784 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1785
1786 if (LV) {
1787 // Update live variables
1788 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1789 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1790 if (isKill)
1791 LV->replaceKillInstruction(Src, MI, InsMI);
1792 if (isDead)
1793 LV->replaceKillInstruction(Dest, MI, ExtMI);
1794 }
1795
1796 return ExtMI;
1797}
1798
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]1799/// convertToThreeAddress - This method must be implemented by targets that
1800/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1801/// may be able to convert a two-address instruction into a true
1802/// three-address instruction on demand. This allows the X86 target (for
1803/// example) to convert ADD and SHL instructions into LEA instructions if they
1804/// would require register copies due to two-addressness.
1805///
1806/// This method returns a null pointer if the transformation cannot be
1807/// performed, otherwise it returns the new instruction.
1808///
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1809MachineInstr *
1810X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1811 MachineBasicBlock::iterator &MBBI,
Owen Andersonf660c172008-07-02 23:41:07 +0000[diff] [blame]1812 LiveVariables *LV) const {
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1813 MachineInstr *MI = MBBI;
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]1814 MachineFunction &MF = *MI->getParent()->getParent();
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]1815 // All instructions input are two-addr instructions. Get the known operands.
1816 unsigned Dest = MI->getOperand(0).getReg();
1817 unsigned Src = MI->getOperand(1).getReg();
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1818 bool isDead = MI->getOperand(0).isDead();
1819 bool isKill = MI->getOperand(1).isKill();
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]1820
Evan Cheng6ce7dc22006-11-15 20:58:11 +0000[diff] [blame]1821 MachineInstr *NewMI = NULL;
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1822 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1823 // we have better subtarget support, enable the 16-bit LEA generation here.
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1824 // 16-bit LEA is also slow on Core2.
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1825 bool DisableLEA16 = true;
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1826 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1827
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1828 unsigned MIOpc = MI->getOpcode();
1829 switch (MIOpc) {
Evan Chengccba76b2006-05-30 20:26:50 +0000[diff] [blame]1830 case X86::SHUFPSrri: {
1831 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1832 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1833
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]1834 unsigned B = MI->getOperand(1).getReg();
1835 unsigned C = MI->getOperand(2).getReg();
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1836 if (B != C) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1837 unsigned A = MI->getOperand(0).getReg();
1838 unsigned M = MI->getOperand(3).getImm();
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1839 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]1840 .addReg(A, RegState::Define | getDeadRegState(isDead))
1841 .addReg(B, getKillRegState(isKill)).addImm(M);
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1842 break;
1843 }
Craig Topper05189702012-01-13 09:21:41 +0000[diff] [blame]1844 case X86::SHUFPDrri: {
1845 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1846 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1847
1848 unsigned B = MI->getOperand(1).getReg();
1849 unsigned C = MI->getOperand(2).getReg();
1850 if (B != C) return 0;
1851 unsigned A = MI->getOperand(0).getReg();
1852 unsigned M = MI->getOperand(3).getImm();
1853
1854 // Convert to PSHUFD mask.
1855 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1856
1857 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1858 .addReg(A, RegState::Define | getDeadRegState(isDead))
1859 .addReg(B, getKillRegState(isKill)).addImm(M);
1860 break;
1861 }
Chris Lattner995f5502007-03-28 18:12:31 +0000[diff] [blame]1862 case X86::SHL64ri: {
Evan Cheng24f2ea32007-09-14 21:48:26 +0000[diff] [blame]1863 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattner995f5502007-03-28 18:12:31 +0000[diff] [blame]1864 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1865 // the flags produced by a shift yet, so this is safe.
Chris Lattner995f5502007-03-28 18:12:31 +0000[diff] [blame]1866 unsigned ShAmt = MI->getOperand(2).getImm();
1867 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1868
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1869 // LEA can't handle RSP.
1870 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1871 !MF.getRegInfo().constrainRegClass(Src, &X86::GR64_NOSPRegClass))
1872 return 0;
1873
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1874 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]1875 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1876 .addReg(0).addImm(1 << ShAmt)
1877 .addReg(Src, getKillRegState(isKill))
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1878 .addImm(0).addReg(0);
Chris Lattner995f5502007-03-28 18:12:31 +0000[diff] [blame]1879 break;
1880 }
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1881 case X86::SHL32ri: {
Evan Cheng24f2ea32007-09-14 21:48:26 +0000[diff] [blame]1882 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1883 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1884 // the flags produced by a shift yet, so this is safe.
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1885 unsigned ShAmt = MI->getOperand(2).getImm();
1886 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1887
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1888 // LEA can't handle ESP.
1889 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1890 !MF.getRegInfo().constrainRegClass(Src, &X86::GR32_NOSPRegClass))
1891 return 0;
1892
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1893 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1894 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]1895 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1896 .addReg(0).addImm(1 << ShAmt)
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1897 .addReg(Src, getKillRegState(isKill)).addImm(0).addReg(0);
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1898 break;
1899 }
1900 case X86::SHL16ri: {
Evan Cheng24f2ea32007-09-14 21:48:26 +0000[diff] [blame]1901 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Evan Cheng61d9c862007-09-06 00:14:41 +0000[diff] [blame]1902 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1903 // the flags produced by a shift yet, so this is safe.
Evan Cheng61d9c862007-09-06 00:14:41 +0000[diff] [blame]1904 unsigned ShAmt = MI->getOperand(2).getImm();
1905 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1906
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1907 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1908 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1909 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1910 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1911 .addReg(0).addImm(1 << ShAmt)
1912 .addReg(Src, getKillRegState(isKill))
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1913 .addImm(0).addReg(0);
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1914 break;
Evan Chengccba76b2006-05-30 20:26:50 +0000[diff] [blame]1915 }
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1916 default: {
1917 // The following opcodes also sets the condition code register(s). Only
1918 // convert them to equivalent lea if the condition code register def's
1919 // are dead!
1920 if (hasLiveCondCodeDef(MI))
1921 return 0;
Evan Chengccba76b2006-05-30 20:26:50 +0000[diff] [blame]1922
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1923 switch (MIOpc) {
1924 default: return 0;
1925 case X86::INC64r:
Dan Gohmancca29832009-01-06 23:34:46 +0000[diff] [blame]1926 case X86::INC32r:
1927 case X86::INC64_32r: {
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1928 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Evan Chengb76143c2007-10-09 07:14:53 +0000[diff] [blame]1929 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1930 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]1931 const TargetRegisterClass *RC = MIOpc == X86::INC64r ?
1932 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1933 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1934
1935 // LEA can't handle RSP.
1936 if (TargetRegisterInfo::isVirtualRegister(Src) &&
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]1937 !MF.getRegInfo().constrainRegClass(Src, RC))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1938 return 0;
1939
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1940 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]1941 .addReg(Dest, RegState::Define |
1942 getDeadRegState(isDead)),
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]1943 Src, isKill, 1);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1944 break;
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]1945 }
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1946 case X86::INC16r:
1947 case X86::INC64_16r:
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1948 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1949 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1950 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1951 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]1952 .addReg(Dest, RegState::Define |
1953 getDeadRegState(isDead)),
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1954 Src, isKill, 1);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1955 break;
1956 case X86::DEC64r:
Dan Gohmancca29832009-01-06 23:34:46 +0000[diff] [blame]1957 case X86::DEC32r:
1958 case X86::DEC64_32r: {
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1959 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Evan Chengb76143c2007-10-09 07:14:53 +0000[diff] [blame]1960 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1961 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]1962 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ?
1963 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1964 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1965 // LEA can't handle RSP.
1966 if (TargetRegisterInfo::isVirtualRegister(Src) &&
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]1967 !MF.getRegInfo().constrainRegClass(Src, RC))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1968 return 0;
1969
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1970 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]1971 .addReg(Dest, RegState::Define |
1972 getDeadRegState(isDead)),
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]1973 Src, isKill, -1);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1974 break;
1975 }
1976 case X86::DEC16r:
1977 case X86::DEC64_16r:
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1978 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1979 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1980 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1981 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]1982 .addReg(Dest, RegState::Define |
1983 getDeadRegState(isDead)),
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1984 Src, isKill, -1);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1985 break;
1986 case X86::ADD64rr:
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]1987 case X86::ADD64rr_DB:
1988 case X86::ADD32rr:
1989 case X86::ADD32rr_DB: {
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1990 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]1991 unsigned Opc;
Craig Topper44d23822012-02-22 05:59:10 +0000[diff] [blame]1992 const TargetRegisterClass *RC;
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]1993 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
1994 Opc = X86::LEA64r;
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]1995 RC = &X86::GR64_NOSPRegClass;
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]1996 } else {
1997 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]1998 RC = &X86::GR32_NOSPRegClass;
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]1999 }
2000
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2001
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2002 unsigned Src2 = MI->getOperand(2).getReg();
2003 bool isKill2 = MI->getOperand(2).isKill();
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]2004
2005 // LEA can't handle RSP.
2006 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]2007 !MF.getRegInfo().constrainRegClass(Src2, RC))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]2008 return 0;
2009
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]2010 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]2011 .addReg(Dest, RegState::Define |
2012 getDeadRegState(isDead)),
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2013 Src, isKill, Src2, isKill2);
2014 if (LV && isKill2)
2015 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2016 break;
2017 }
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]2018 case X86::ADD16rr:
2019 case X86::ADD16rr_DB: {
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2020 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]2021 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2022 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2023 unsigned Src2 = MI->getOperand(2).getReg();
2024 bool isKill2 = MI->getOperand(2).isKill();
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]2025 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]2026 .addReg(Dest, RegState::Define |
2027 getDeadRegState(isDead)),
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2028 Src, isKill, Src2, isKill2);
2029 if (LV && isKill2)
2030 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2031 break;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2032 }
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2033 case X86::ADD64ri32:
2034 case X86::ADD64ri8:
Chris Lattner15df55d2010-10-08 03:57:25 +0000[diff] [blame]2035 case X86::ADD64ri32_DB:
2036 case X86::ADD64ri8_DB:
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2037 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]2038 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2039 .addReg(Dest, RegState::Define |
2040 getDeadRegState(isDead)),
2041 Src, isKill, MI->getOperand(2).getImm());
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2042 break;
2043 case X86::ADD32ri:
Chris Lattner15df55d2010-10-08 03:57:25 +0000[diff] [blame]2044 case X86::ADD32ri8:
2045 case X86::ADD32ri_DB:
2046 case X86::ADD32ri8_DB: {
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2047 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2048 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]2049 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2050 .addReg(Dest, RegState::Define |
2051 getDeadRegState(isDead)),
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]2052 Src, isKill, MI->getOperand(2).getImm());
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2053 break;
2054 }
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2055 case X86::ADD16ri:
2056 case X86::ADD16ri8:
Chris Lattner15df55d2010-10-08 03:57:25 +0000[diff] [blame]2057 case X86::ADD16ri_DB:
2058 case X86::ADD16ri8_DB:
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2059 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]2060 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2061 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]2062 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2063 .addReg(Dest, RegState::Define |
2064 getDeadRegState(isDead)),
2065 Src, isKill, MI->getOperand(2).getImm());
2066 break;
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2067 }
2068 }
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]2069 }
2070
Evan Cheng15246732008-02-07 08:29:53 +0000[diff] [blame]2071 if (!NewMI) return 0;
2072
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2073 if (LV) { // Update live variables
2074 if (isKill)
2075 LV->replaceKillInstruction(Src, MI, NewMI);
2076 if (isDead)
2077 LV->replaceKillInstruction(Dest, MI, NewMI);
2078 }
2079
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2080 MFI->insert(MBBI, NewMI); // Insert the new inst
Evan Cheng6ce7dc22006-11-15 20:58:11 +0000[diff] [blame]2081 return NewMI;
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]2082}
2083
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2084/// commuteInstruction - We have a few instructions that must be hacked on to
2085/// commute them.
2086///
Evan Cheng58dcb0e2008-06-16 07:33:11 +0000[diff] [blame]2087MachineInstr *
2088X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2089 switch (MI->getOpcode()) {
Chris Lattner0df53d22005-01-19 07:31:24 +0000[diff] [blame]2090 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2091 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2092 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
Dan Gohmane47f1f92007-09-14 23:17:45 +0000[diff] [blame]2093 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2094 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2095 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
Chris Lattner0df53d22005-01-19 07:31:24 +0000[diff] [blame]2096 unsigned Opc;
2097 unsigned Size;
2098 switch (MI->getOpcode()) {
Torok Edwinc23197a2009-07-14 16:55:14 +0000[diff] [blame]2099 default: llvm_unreachable("Unreachable!");
Chris Lattner0df53d22005-01-19 07:31:24 +0000[diff] [blame]2100 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2101 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2102 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2103 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
Dan Gohmane47f1f92007-09-14 23:17:45 +0000[diff] [blame]2104 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2105 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
Chris Lattner0df53d22005-01-19 07:31:24 +0000[diff] [blame]2106 }
Chris Lattner9a1ceae2007-12-30 20:49:49 +0000[diff] [blame]2107 unsigned Amt = MI->getOperand(3).getImm();
Dan Gohman74feef22008-10-17 01:23:35 +0000[diff] [blame]2108 if (NewMI) {
2109 MachineFunction &MF = *MI->getParent()->getParent();
2110 MI = MF.CloneMachineInstr(MI);
2111 NewMI = false;
Evan Chenga4d16a12008-02-13 02:46:49 +0000[diff] [blame]2112 }
Dan Gohman74feef22008-10-17 01:23:35 +0000[diff] [blame]2113 MI->setDesc(get(Opc));
2114 MI->getOperand(3).setImm(Size-Amt);
2115 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2116 }
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2117 case X86::CMOVB16rr:
2118 case X86::CMOVB32rr:
2119 case X86::CMOVB64rr:
2120 case X86::CMOVAE16rr:
2121 case X86::CMOVAE32rr:
2122 case X86::CMOVAE64rr:
2123 case X86::CMOVE16rr:
2124 case X86::CMOVE32rr:
2125 case X86::CMOVE64rr:
2126 case X86::CMOVNE16rr:
2127 case X86::CMOVNE32rr:
2128 case X86::CMOVNE64rr:
Chris Lattner25cbf502010-10-05 23:00:14 +0000[diff] [blame]2129 case X86::CMOVBE16rr:
2130 case X86::CMOVBE32rr:
2131 case X86::CMOVBE64rr:
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2132 case X86::CMOVA16rr:
2133 case X86::CMOVA32rr:
2134 case X86::CMOVA64rr:
2135 case X86::CMOVL16rr:
2136 case X86::CMOVL32rr:
2137 case X86::CMOVL64rr:
2138 case X86::CMOVGE16rr:
2139 case X86::CMOVGE32rr:
2140 case X86::CMOVGE64rr:
2141 case X86::CMOVLE16rr:
2142 case X86::CMOVLE32rr:
2143 case X86::CMOVLE64rr:
2144 case X86::CMOVG16rr:
2145 case X86::CMOVG32rr:
2146 case X86::CMOVG64rr:
2147 case X86::CMOVS16rr:
2148 case X86::CMOVS32rr:
2149 case X86::CMOVS64rr:
2150 case X86::CMOVNS16rr:
2151 case X86::CMOVNS32rr:
2152 case X86::CMOVNS64rr:
2153 case X86::CMOVP16rr:
2154 case X86::CMOVP32rr:
2155 case X86::CMOVP64rr:
2156 case X86::CMOVNP16rr:
2157 case X86::CMOVNP32rr:
Dan Gohman305fceb2009-01-07 00:35:10 +0000[diff] [blame]2158 case X86::CMOVNP64rr:
2159 case X86::CMOVO16rr:
2160 case X86::CMOVO32rr:
2161 case X86::CMOVO64rr:
2162 case X86::CMOVNO16rr:
2163 case X86::CMOVNO32rr:
2164 case X86::CMOVNO64rr: {
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2165 unsigned Opc = 0;
2166 switch (MI->getOpcode()) {
2167 default: break;
2168 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2169 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2170 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2171 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2172 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2173 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2174 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2175 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2176 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2177 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2178 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2179 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
Chris Lattner25cbf502010-10-05 23:00:14 +0000[diff] [blame]2180 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2181 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2182 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2183 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2184 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2185 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2186 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2187 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2188 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2189 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2190 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2191 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2192 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2193 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2194 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2195 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2196 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2197 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2198 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2199 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
Mon P Wang0bd07fc2009-04-18 05:16:01 +0000[diff] [blame]2200 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2201 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2202 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2203 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2204 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2205 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
Mon P Wang0bd07fc2009-04-18 05:16:01 +0000[diff] [blame]2206 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2207 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2208 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2209 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
Dan Gohman305fceb2009-01-07 00:35:10 +0000[diff] [blame]2210 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2211 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
Mon P Wang0bd07fc2009-04-18 05:16:01 +0000[diff] [blame]2212 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
Dan Gohman305fceb2009-01-07 00:35:10 +0000[diff] [blame]2213 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2214 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2215 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2216 }
Dan Gohman74feef22008-10-17 01:23:35 +0000[diff] [blame]2217 if (NewMI) {
2218 MachineFunction &MF = *MI->getParent()->getParent();
2219 MI = MF.CloneMachineInstr(MI);
2220 NewMI = false;
2221 }
Chris Lattner5080f4d2008-01-11 18:10:50 +0000[diff] [blame]2222 MI->setDesc(get(Opc));
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2223 // Fallthrough intended.
2224 }
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2225 default:
Evan Cheng58dcb0e2008-06-16 07:33:11 +0000[diff] [blame]2226 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2227 }
2228}
2229
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2230static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
2231 switch (BrOpc) {
2232 default: return X86::COND_INVALID;
Chris Lattnerbd13fb62010-02-11 19:25:55 +0000[diff] [blame]2233 case X86::JE_4: return X86::COND_E;
2234 case X86::JNE_4: return X86::COND_NE;
2235 case X86::JL_4: return X86::COND_L;
2236 case X86::JLE_4: return X86::COND_LE;
2237 case X86::JG_4: return X86::COND_G;
2238 case X86::JGE_4: return X86::COND_GE;
2239 case X86::JB_4: return X86::COND_B;
2240 case X86::JBE_4: return X86::COND_BE;
2241 case X86::JA_4: return X86::COND_A;
2242 case X86::JAE_4: return X86::COND_AE;
2243 case X86::JS_4: return X86::COND_S;
2244 case X86::JNS_4: return X86::COND_NS;
2245 case X86::JP_4: return X86::COND_P;
2246 case X86::JNP_4: return X86::COND_NP;
2247 case X86::JO_4: return X86::COND_O;
2248 case X86::JNO_4: return X86::COND_NO;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2249 }
2250}
2251
2252unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2253 switch (CC) {
Torok Edwinc23197a2009-07-14 16:55:14 +0000[diff] [blame]2254 default: llvm_unreachable("Illegal condition code!");
Chris Lattnerbd13fb62010-02-11 19:25:55 +0000[diff] [blame]2255 case X86::COND_E: return X86::JE_4;
2256 case X86::COND_NE: return X86::JNE_4;
2257 case X86::COND_L: return X86::JL_4;
2258 case X86::COND_LE: return X86::JLE_4;
2259 case X86::COND_G: return X86::JG_4;
2260 case X86::COND_GE: return X86::JGE_4;
2261 case X86::COND_B: return X86::JB_4;
2262 case X86::COND_BE: return X86::JBE_4;
2263 case X86::COND_A: return X86::JA_4;
2264 case X86::COND_AE: return X86::JAE_4;
2265 case X86::COND_S: return X86::JS_4;
2266 case X86::COND_NS: return X86::JNS_4;
2267 case X86::COND_P: return X86::JP_4;
2268 case X86::COND_NP: return X86::JNP_4;
2269 case X86::COND_O: return X86::JO_4;
2270 case X86::COND_NO: return X86::JNO_4;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2271 }
2272}
2273
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]2274/// GetOppositeBranchCondition - Return the inverse of the specified condition,
2275/// e.g. turning COND_E to COND_NE.
2276X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2277 switch (CC) {
Torok Edwinc23197a2009-07-14 16:55:14 +0000[diff] [blame]2278 default: llvm_unreachable("Illegal condition code!");
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]2279 case X86::COND_E: return X86::COND_NE;
2280 case X86::COND_NE: return X86::COND_E;
2281 case X86::COND_L: return X86::COND_GE;
2282 case X86::COND_LE: return X86::COND_G;
2283 case X86::COND_G: return X86::COND_LE;
2284 case X86::COND_GE: return X86::COND_L;
2285 case X86::COND_B: return X86::COND_AE;
2286 case X86::COND_BE: return X86::COND_A;
2287 case X86::COND_A: return X86::COND_BE;
2288 case X86::COND_AE: return X86::COND_B;
2289 case X86::COND_S: return X86::COND_NS;
2290 case X86::COND_NS: return X86::COND_S;
2291 case X86::COND_P: return X86::COND_NP;
2292 case X86::COND_NP: return X86::COND_P;
2293 case X86::COND_O: return X86::COND_NO;
2294 case X86::COND_NO: return X86::COND_O;
2295 }
2296}
2297
Jakob Stoklund Olesen59bde4d2012-07-04 00:09:58 +0000[diff] [blame]2298/// getCMovFromCond - Return a cmov(rr) opcode for the given condition and
2299/// register size in bytes.
2300static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes) {
2301 static const unsigned Opc[16][3] = {
2302 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2303 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2304 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2305 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2306 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2307 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2308 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2309 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2310 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2311 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2312 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2313 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2314 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2315 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2316 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
2317 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr }
2318 };
2319
2320 assert(CC < 16 && "Can only handle standard cond codes");
2321 switch(RegBytes) {
2322 default: llvm_unreachable("Illegal register size!");
2323 case 2: return Opc[CC][0];
2324 case 4: return Opc[CC][1];
2325 case 8: return Opc[CC][2];
2326 }
2327}
2328
Dale Johannesen318093b2007-06-14 22:03:45 +0000[diff] [blame]2329bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
Evan Cheng5a96b3d2011-12-07 07:15:52 +0000[diff] [blame]2330 if (!MI->isTerminator()) return false;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2331
Chris Lattner69244302008-01-07 01:56:04 +0000[diff] [blame]2332 // Conditional branch is a special case.
Evan Cheng5a96b3d2011-12-07 07:15:52 +0000[diff] [blame]2333 if (MI->isBranch() && !MI->isBarrier())
Chris Lattner69244302008-01-07 01:56:04 +0000[diff] [blame]2334 return true;
Evan Cheng5a96b3d2011-12-07 07:15:52 +0000[diff] [blame]2335 if (!MI->isPredicable())
Chris Lattner69244302008-01-07 01:56:04 +0000[diff] [blame]2336 return true;
2337 return !isPredicated(MI);
Dale Johannesen318093b2007-06-14 22:03:45 +0000[diff] [blame]2338}
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]2339
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2340bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2341 MachineBasicBlock *&TBB,
2342 MachineBasicBlock *&FBB,
Evan Chengdc54d312009-02-09 07:14:22 +0000[diff] [blame]2343 SmallVectorImpl<MachineOperand> &Cond,
2344 bool AllowModify) const {
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2345 // Start from the bottom of the block and work up, examining the
2346 // terminator instructions.
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2347 MachineBasicBlock::iterator I = MBB.end();
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2348 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2349 while (I != MBB.begin()) {
2350 --I;
Dale Johannesen93d6a7e2010-04-02 01:38:09 +0000[diff] [blame]2351 if (I->isDebugValue())
2352 continue;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2353
2354 // Working from the bottom, when we see a non-terminator instruction, we're
2355 // done.
Jakob Stoklund Olesen468a2a42010-07-16 17:41:44 +0000[diff] [blame]2356 if (!isUnpredicatedTerminator(I))
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2357 break;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2358
2359 // A terminator that isn't a branch can't easily be handled by this
2360 // analysis.
Evan Cheng5a96b3d2011-12-07 07:15:52 +0000[diff] [blame]2361 if (!I->isBranch())
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2362 return true;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2363
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2364 // Handle unconditional branches.
Chris Lattnerbd13fb62010-02-11 19:25:55 +0000[diff] [blame]2365 if (I->getOpcode() == X86::JMP_4) {
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2366 UnCondBrIter = I;
2367
Evan Chengdc54d312009-02-09 07:14:22 +0000[diff] [blame]2368 if (!AllowModify) {
2369 TBB = I->getOperand(0).getMBB();
Evan Cheng45e00102009-05-08 06:34:09 +0000[diff] [blame]2370 continue;
Evan Chengdc54d312009-02-09 07:14:22 +0000[diff] [blame]2371 }
2372
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2373 // If the block has any instructions after a JMP, delete them.
Chris Lattner7896c9f2009-12-03 00:50:42 +0000[diff] [blame]2374 while (llvm::next(I) != MBB.end())
2375 llvm::next(I)->eraseFromParent();
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2376
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2377 Cond.clear();
2378 FBB = 0;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2379
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2380 // Delete the JMP if it's equivalent to a fall-through.
2381 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2382 TBB = 0;
2383 I->eraseFromParent();
2384 I = MBB.end();
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2385 UnCondBrIter = MBB.end();
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2386 continue;
2387 }
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2388
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2389 // TBB is used to indicate the unconditional destination.
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2390 TBB = I->getOperand(0).getMBB();
2391 continue;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2392 }
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2393
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2394 // Handle conditional branches.
2395 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2396 if (BranchCode == X86::COND_INVALID)
2397 return true; // Can't handle indirect branch.
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2398
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2399 // Working from the bottom, handle the first conditional branch.
2400 if (Cond.empty()) {
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2401 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2402 if (AllowModify && UnCondBrIter != MBB.end() &&
2403 MBB.isLayoutSuccessor(TargetBB)) {
2404 // If we can modify the code and it ends in something like:
2405 //
2406 // jCC L1
2407 // jmp L2
2408 // L1:
2409 // ...
2410 // L2:
2411 //
2412 // Then we can change this to:
2413 //
2414 // jnCC L2
2415 // L1:
2416 // ...
2417 // L2:
2418 //
2419 // Which is a bit more efficient.
2420 // We conditionally jump to the fall-through block.
2421 BranchCode = GetOppositeBranchCondition(BranchCode);
2422 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2423 MachineBasicBlock::iterator OldInst = I;
2424
2425 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2426 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2427 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2428 .addMBB(TargetBB);
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2429
2430 OldInst->eraseFromParent();
2431 UnCondBrIter->eraseFromParent();
2432
2433 // Restart the analysis.
2434 UnCondBrIter = MBB.end();
2435 I = MBB.end();
2436 continue;
2437 }
2438
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2439 FBB = TBB;
2440 TBB = I->getOperand(0).getMBB();
2441 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2442 continue;
2443 }
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2444
2445 // Handle subsequent conditional branches. Only handle the case where all
2446 // conditional branches branch to the same destination and their condition
2447 // opcodes fit one of the special multi-branch idioms.
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2448 assert(Cond.size() == 1);
2449 assert(TBB);
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2450
2451 // Only handle the case where all conditional branches branch to the same
2452 // destination.
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2453 if (TBB != I->getOperand(0).getMBB())
2454 return true;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2455
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2456 // If the conditions are the same, we can leave them alone.
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2457 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2458 if (OldBranchCode == BranchCode)
2459 continue;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2460
2461 // If they differ, see if they fit one of the known patterns. Theoretically,
2462 // we could handle more patterns here, but we shouldn't expect to see them
2463 // if instruction selection has done a reasonable job.
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2464 if ((OldBranchCode == X86::COND_NP &&
2465 BranchCode == X86::COND_E) ||
2466 (OldBranchCode == X86::COND_E &&
2467 BranchCode == X86::COND_NP))
2468 BranchCode = X86::COND_NP_OR_E;
2469 else if ((OldBranchCode == X86::COND_P &&
2470 BranchCode == X86::COND_NE) ||
2471 (OldBranchCode == X86::COND_NE &&
2472 BranchCode == X86::COND_P))
2473 BranchCode = X86::COND_NE_OR_P;
2474 else
2475 return true;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2476
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2477 // Update the MachineOperand.
2478 Cond[0].setImm(BranchCode);
Chris Lattner6ce64432006-10-30 22:27:23 +0000[diff] [blame]2479 }
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2480
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2481 return false;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2482}
2483
Evan Cheng6ae36262007-05-18 00:18:17 +0000[diff] [blame]2484unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2485 MachineBasicBlock::iterator I = MBB.end();
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2486 unsigned Count = 0;
2487
2488 while (I != MBB.begin()) {
2489 --I;
Dale Johannesen93d6a7e2010-04-02 01:38:09 +0000[diff] [blame]2490 if (I->isDebugValue())
2491 continue;
Chris Lattnerbd13fb62010-02-11 19:25:55 +0000[diff] [blame]2492 if (I->getOpcode() != X86::JMP_4 &&
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2493 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2494 break;
2495 // Remove the branch.
2496 I->eraseFromParent();
2497 I = MBB.end();
2498 ++Count;
2499 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2500
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2501 return Count;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2502}
2503
Evan Cheng6ae36262007-05-18 00:18:17 +0000[diff] [blame]2504unsigned
2505X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2506 MachineBasicBlock *FBB,
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2507 const SmallVectorImpl<MachineOperand> &Cond,
2508 DebugLoc DL) const {
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2509 // Shouldn't be a fall through.
2510 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
Chris Lattner34a84ac2006-10-21 05:34:23 +0000[diff] [blame]2511 assert((Cond.size() == 1 || Cond.size() == 0) &&
2512 "X86 branch conditions have one component!");
2513
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2514 if (Cond.empty()) {
2515 // Unconditional branch?
2516 assert(!FBB && "Unconditional branch with multiple successors!");
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2517 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
Evan Cheng6ae36262007-05-18 00:18:17 +0000[diff] [blame]2518 return 1;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2519 }
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2520
2521 // Conditional branch.
2522 unsigned Count = 0;
2523 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2524 switch (CC) {
2525 case X86::COND_NP_OR_E:
2526 // Synthesize NP_OR_E with two branches.
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2527 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2528 ++Count;
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2529 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2530 ++Count;
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2531 break;
2532 case X86::COND_NE_OR_P:
2533 // Synthesize NE_OR_P with two branches.
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2534 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2535 ++Count;
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2536 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2537 ++Count;
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2538 break;
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2539 default: {
2540 unsigned Opc = GetCondBranchFromCond(CC);
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2541 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2542 ++Count;
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2543 }
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2544 }
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2545 if (FBB) {
2546 // Two-way Conditional branch. Insert the second branch.
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2547 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2548 ++Count;
2549 }
2550 return Count;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2551}
2552
Jakob Stoklund Olesen59bde4d2012-07-04 00:09:58 +0000[diff] [blame]2553bool X86InstrInfo::
2554canInsertSelect(const MachineBasicBlock &MBB,
2555 const SmallVectorImpl<MachineOperand> &Cond,
2556 unsigned TrueReg, unsigned FalseReg,
2557 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2558 // Not all subtargets have cmov instructions.
2559 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2560 return false;
2561 if (Cond.size() != 1)
2562 return false;
2563 // We cannot do the composite conditions, at least not in SSA form.
2564 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2565 return false;
2566
2567 // Check register classes.
2568 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2569 const TargetRegisterClass *RC =
2570 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2571 if (!RC)
2572 return false;
2573
2574 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2575 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2576 X86::GR32RegClass.hasSubClassEq(RC) ||
2577 X86::GR64RegClass.hasSubClassEq(RC)) {
2578 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2579 // Bridge. Probably Ivy Bridge as well.
2580 CondCycles = 2;
2581 TrueCycles = 2;
2582 FalseCycles = 2;
2583 return true;
2584 }
2585
2586 // Can't do vectors.
2587 return false;
2588}
2589
2590void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2591 MachineBasicBlock::iterator I, DebugLoc DL,
2592 unsigned DstReg,
2593 const SmallVectorImpl<MachineOperand> &Cond,
2594 unsigned TrueReg, unsigned FalseReg) const {
2595 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2596 assert(Cond.size() == 1 && "Invalid Cond array");
2597 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
2598 MRI.getRegClass(DstReg)->getSize());
2599 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2600}
2601
Dan Gohman6d9305c2009-04-15 00:04:23 +0000[diff] [blame]2602/// isHReg - Test if the given register is a physical h register.
2603static bool isHReg(unsigned Reg) {
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2604 return X86::GR8_ABCD_HRegClass.contains(Reg);
Dan Gohman6d9305c2009-04-15 00:04:23 +0000[diff] [blame]2605}
2606
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2607// Try and copy between VR128/VR64 and GR64 registers.
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2608static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2609 bool HasAVX) {
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2610 // SrcReg(VR128) -> DestReg(GR64)
2611 // SrcReg(VR64) -> DestReg(GR64)
2612 // SrcReg(GR64) -> DestReg(VR128)
2613 // SrcReg(GR64) -> DestReg(VR64)
2614
2615 if (X86::GR64RegClass.contains(DestReg)) {
2616 if (X86::VR128RegClass.contains(SrcReg)) {
2617 // Copy from a VR128 register to a GR64 register.
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2618 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2619 } else if (X86::VR64RegClass.contains(SrcReg)) {
2620 // Copy from a VR64 register to a GR64 register.
2621 return X86::MOVSDto64rr;
2622 }
2623 } else if (X86::GR64RegClass.contains(SrcReg)) {
2624 // Copy from a GR64 register to a VR128 register.
2625 if (X86::VR128RegClass.contains(DestReg))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2626 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2627 // Copy from a GR64 register to a VR64 register.
2628 else if (X86::VR64RegClass.contains(DestReg))
2629 return X86::MOV64toSDrr;
2630 }
2631
Jakob Stoklund Olesen4bd89872011-09-22 22:45:24 +0000[diff] [blame]2632 // SrcReg(FR32) -> DestReg(GR32)
2633 // SrcReg(GR32) -> DestReg(FR32)
2634
2635 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
2636 // Copy from a FR32 register to a GR32 register.
2637 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
2638
2639 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
2640 // Copy from a GR32 register to a FR32 register.
2641 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
2642
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2643 return 0;
2644}
2645
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2646void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2647 MachineBasicBlock::iterator MI, DebugLoc DL,
2648 unsigned DestReg, unsigned SrcReg,
2649 bool KillSrc) const {
2650 // First deal with the normal symmetric copies.
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2651 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2652 unsigned Opc = 0;
2653 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2654 Opc = X86::MOV64rr;
2655 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2656 Opc = X86::MOV32rr;
2657 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2658 Opc = X86::MOV16rr;
2659 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2660 // Copying to or from a physical H register on x86-64 requires a NOREX
2661 // move. Otherwise use a normal move.
2662 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
Jakob Stoklund Olesenb66f1842011-10-07 20:15:54 +0000[diff] [blame]2663 TM.getSubtarget<X86Subtarget>().is64Bit()) {
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2664 Opc = X86::MOV8rr_NOREX;
Jakob Stoklund Olesenb66f1842011-10-07 20:15:54 +0000[diff] [blame]2665 // Both operands must be encodable without an REX prefix.
2666 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2667 "8-bit H register can not be copied outside GR8_NOREX");
2668 } else
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2669 Opc = X86::MOV8rr;
2670 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2671 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]2672 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2673 Opc = X86::VMOVAPSYrr;
Jakob Stoklund Olesen61c8ecc2010-07-08 22:30:35 +0000[diff] [blame]2674 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2675 Opc = X86::MMX_MOVQ64rr;
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2676 else
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2677 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2678
2679 if (Opc) {
2680 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2681 .addReg(SrcReg, getKillRegState(KillSrc));
2682 return;
2683 }
2684
2685 // Moving EFLAGS to / from another register requires a push and a pop.
2686 if (SrcReg == X86::EFLAGS) {
2687 if (X86::GR64RegClass.contains(DestReg)) {
2688 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2689 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2690 return;
2691 } else if (X86::GR32RegClass.contains(DestReg)) {
2692 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2693 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2694 return;
2695 }
2696 }
2697 if (DestReg == X86::EFLAGS) {
2698 if (X86::GR64RegClass.contains(SrcReg)) {
2699 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2700 .addReg(SrcReg, getKillRegState(KillSrc));
2701 BuildMI(MBB, MI, DL, get(X86::POPF64));
2702 return;
2703 } else if (X86::GR32RegClass.contains(SrcReg)) {
2704 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2705 .addReg(SrcReg, getKillRegState(KillSrc));
2706 BuildMI(MBB, MI, DL, get(X86::POPF32));
2707 return;
2708 }
2709 }
2710
2711 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2712 << " to " << RI.getName(DestReg) << '\n');
2713 llvm_unreachable("Cannot emit physreg copy instruction");
2714}
2715
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2716static unsigned getLoadStoreRegOpcode(unsigned Reg,
2717 const TargetRegisterClass *RC,
2718 bool isStackAligned,
2719 const TargetMachine &TM,
2720 bool load) {
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2721 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2722 switch (RC->getSize()) {
Rafael Espindola5a717a32010-07-12 03:43:04 +0000[diff] [blame]2723 default:
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2724 llvm_unreachable("Unknown spill size");
2725 case 1:
2726 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2727 if (TM.getSubtarget<X86Subtarget>().is64Bit())
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2728 // Copying to or from a physical H register on x86-64 requires a NOREX
2729 // move. Otherwise use a normal move.
2730 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2731 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2732 return load ? X86::MOV8rm : X86::MOV8mr;
2733 case 2:
2734 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2735 return load ? X86::MOV16rm : X86::MOV16mr;
2736 case 4:
2737 if (X86::GR32RegClass.hasSubClassEq(RC))
2738 return load ? X86::MOV32rm : X86::MOV32mr;
2739 if (X86::FR32RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2740 return load ?
2741 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2742 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2743 if (X86::RFP32RegClass.hasSubClassEq(RC))
2744 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2745 llvm_unreachable("Unknown 4-byte regclass");
2746 case 8:
2747 if (X86::GR64RegClass.hasSubClassEq(RC))
2748 return load ? X86::MOV64rm : X86::MOV64mr;
2749 if (X86::FR64RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2750 return load ?
2751 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
2752 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2753 if (X86::VR64RegClass.hasSubClassEq(RC))
2754 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2755 if (X86::RFP64RegClass.hasSubClassEq(RC))
2756 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2757 llvm_unreachable("Unknown 8-byte regclass");
2758 case 10:
2759 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2760 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
Bruno Cardoso Lopes5affa512011-08-31 03:04:09 +0000[diff] [blame]2761 case 16: {
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2762 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2763 // If stack is realigned we can use aligned stores.
2764 if (isStackAligned)
Bruno Cardoso Lopes5affa512011-08-31 03:04:09 +0000[diff] [blame]2765 return load ?
2766 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
2767 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2768 else
Bruno Cardoso Lopes5affa512011-08-31 03:04:09 +0000[diff] [blame]2769 return load ?
2770 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
2771 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
2772 }
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]2773 case 32:
2774 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2775 // If stack is realigned we can use aligned stores.
2776 if (isStackAligned)
2777 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2778 else
2779 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2780 }
2781}
2782
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2783static unsigned getStoreRegOpcode(unsigned SrcReg,
2784 const TargetRegisterClass *RC,
2785 bool isStackAligned,
2786 TargetMachine &TM) {
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2787 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2788}
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2789
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2790
2791static unsigned getLoadRegOpcode(unsigned DestReg,
2792 const TargetRegisterClass *RC,
2793 bool isStackAligned,
2794 const TargetMachine &TM) {
2795 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2796}
2797
2798void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
2799 MachineBasicBlock::iterator MI,
2800 unsigned SrcReg, bool isKill, int FrameIdx,
Evan Cheng746ad692010-05-06 19:06:44 +0000[diff] [blame]2801 const TargetRegisterClass *RC,
2802 const TargetRegisterInfo *TRI) const {
Anton Korobeynikov88bbf692008-07-19 06:30:51 +0000[diff] [blame]2803 const MachineFunction &MF = *MBB.getParent();
Jakob Stoklund Olesen516cd452010-07-27 04:16:58 +0000[diff] [blame]2804 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
2805 "Stack slot too small for store");
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2806 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2807 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Cheng2fa82bc2011-06-23 01:53:43 +0000[diff] [blame]2808 RI.canRealignStack(MF);
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2809 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Dale Johannesen6ec25f52010-01-26 00:03:12 +0000[diff] [blame]2810 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]2811 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]2812 .addReg(SrcReg, getKillRegState(isKill));
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2813}
2814
2815void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
2816 bool isKill,
2817 SmallVectorImpl<MachineOperand> &Addr,
2818 const TargetRegisterClass *RC,
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]2819 MachineInstr::mmo_iterator MMOBegin,
2820 MachineInstr::mmo_iterator MMOEnd,
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2821 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2822 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2823 bool isAligned = MMOBegin != MMOEnd &&
2824 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2825 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Chris Lattnerc7f3ace2010-04-02 20:16:16 +0000[diff] [blame]2826 DebugLoc DL;
Dale Johannesen21b55412009-02-12 23:08:38 +0000[diff] [blame]2827 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2828 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]2829 MIB.addOperand(Addr[i]);
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]2830 MIB.addReg(SrcReg, getKillRegState(isKill));
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]2831 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2832 NewMIs.push_back(MIB);
2833}
2834
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2835
2836void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
Anton Korobeynikov88bbf692008-07-19 06:30:51 +0000[diff] [blame]2837 MachineBasicBlock::iterator MI,
2838 unsigned DestReg, int FrameIdx,
Evan Cheng746ad692010-05-06 19:06:44 +0000[diff] [blame]2839 const TargetRegisterClass *RC,
2840 const TargetRegisterInfo *TRI) const {
Anton Korobeynikov88bbf692008-07-19 06:30:51 +0000[diff] [blame]2841 const MachineFunction &MF = *MBB.getParent();
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2842 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2843 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Cheng2fa82bc2011-06-23 01:53:43 +0000[diff] [blame]2844 RI.canRealignStack(MF);
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2845 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Dale Johannesen6ec25f52010-01-26 00:03:12 +0000[diff] [blame]2846 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]2847 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2848}
2849
2850void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2851 SmallVectorImpl<MachineOperand> &Addr,
2852 const TargetRegisterClass *RC,
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]2853 MachineInstr::mmo_iterator MMOBegin,
2854 MachineInstr::mmo_iterator MMOEnd,
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2855 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2856 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2857 bool isAligned = MMOBegin != MMOEnd &&
2858 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2859 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Chris Lattnerc7f3ace2010-04-02 20:16:16 +0000[diff] [blame]2860 DebugLoc DL;
Dale Johannesen21b55412009-02-12 23:08:38 +0000[diff] [blame]2861 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2862 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]2863 MIB.addOperand(Addr[i]);
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]2864 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2865 NewMIs.push_back(MIB);
2866}
2867
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]2868bool X86InstrInfo::
2869analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
2870 int &CmpMask, int &CmpValue) const {
2871 switch (MI->getOpcode()) {
2872 default: break;
2873 case X86::CMP64ri32:
2874 case X86::CMP64ri8:
2875 case X86::CMP32ri:
2876 case X86::CMP32ri8:
2877 case X86::CMP16ri:
2878 case X86::CMP16ri8:
2879 case X86::CMP8ri:
2880 SrcReg = MI->getOperand(0).getReg();
2881 SrcReg2 = 0;
2882 CmpMask = ~0;
2883 CmpValue = MI->getOperand(1).getImm();
2884 return true;
2885 case X86::CMP64rr:
2886 case X86::CMP32rr:
2887 case X86::CMP16rr:
2888 case X86::CMP8rr:
2889 SrcReg = MI->getOperand(0).getReg();
2890 SrcReg2 = MI->getOperand(1).getReg();
2891 CmpMask = ~0;
2892 CmpValue = 0;
2893 return true;
2894 }
2895 return false;
2896}
2897
2898/// getSwappedConditionForSET - assume the flags are set by MI(a,b), return
2899/// the opcode if we modify the instructions such that flags are
2900/// set by MI(b,a).
2901static unsigned getSwappedConditionForSET(unsigned SETOpc) {
2902 switch (SETOpc) {
2903 default: return 0;
2904 case X86::SETEr: return X86::SETEr;
2905 case X86::SETEm: return X86::SETEm;
2906 case X86::SETNEr: return X86::SETNEr;
2907 case X86::SETNEm: return X86::SETNEm;
2908 case X86::SETLr: return X86::SETGr;
2909 case X86::SETLm: return X86::SETGm;
2910 case X86::SETLEr: return X86::SETGEr;
2911 case X86::SETLEm: return X86::SETGEm;
2912 case X86::SETGr: return X86::SETLr;
2913 case X86::SETGm: return X86::SETLm;
2914 case X86::SETGEr: return X86::SETLEr;
2915 case X86::SETGEm: return X86::SETLEm;
2916 case X86::SETBr: return X86::SETAr;
2917 case X86::SETBm: return X86::SETAm;
2918 case X86::SETBEr: return X86::SETAEr;
2919 case X86::SETBEm: return X86::SETAEm;
2920 case X86::SETAr: return X86::SETBr;
2921 case X86::SETAm: return X86::SETBm;
2922 case X86::SETAEr: return X86::SETBEr;
2923 case X86::SETAEm: return X86::SETBEm;
2924 }
2925}
2926
2927/// getSwappedConditionForBranch - assume the flags are set by MI(a,b), return
2928/// the opcode if we modify the instructions such that flags are
2929/// set by MI(b,a).
2930static unsigned getSwappedConditionForBranch(unsigned BranchOpc) {
2931 switch (BranchOpc) {
2932 default: return 0;
2933 case X86::JE_4: return X86::JE_4;
2934 case X86::JNE_4: return X86::JNE_4;
2935 case X86::JL_4: return X86::JG_4;
2936 case X86::JLE_4: return X86::JGE_4;
2937 case X86::JG_4: return X86::JL_4;
2938 case X86::JGE_4: return X86::JLE_4;
2939 case X86::JB_4: return X86::JA_4;
2940 case X86::JBE_4: return X86::JAE_4;
2941 case X86::JA_4: return X86::JB_4;
2942 case X86::JAE_4: return X86::JBE_4;
2943 }
2944}
2945
2946/// getSwappedConditionForCMov - assume the flags are set by MI(a,b), return
2947/// the opcode if we modify the instructions such that flags are
2948/// set by MI(b,a).
2949static unsigned getSwappedConditionForCMov(unsigned CMovOpc) {
2950 switch (CMovOpc) {
2951 default: return 0;
2952 case X86::CMOVE16rm: return X86::CMOVE16rm;
2953 case X86::CMOVE16rr: return X86::CMOVE16rr;
2954 case X86::CMOVE32rm: return X86::CMOVE32rm;
2955 case X86::CMOVE32rr: return X86::CMOVE32rr;
2956 case X86::CMOVE64rm: return X86::CMOVE64rm;
2957 case X86::CMOVE64rr: return X86::CMOVE64rr;
2958 case X86::CMOVNE16rm: return X86::CMOVNE16rm;
2959 case X86::CMOVNE16rr: return X86::CMOVNE16rr;
2960 case X86::CMOVNE32rm: return X86::CMOVNE32rm;
2961 case X86::CMOVNE32rr: return X86::CMOVNE32rr;
2962 case X86::CMOVNE64rm: return X86::CMOVNE64rm;
2963 case X86::CMOVNE64rr: return X86::CMOVNE64rr;
2964
2965 case X86::CMOVL16rm: return X86::CMOVG16rm;
2966 case X86::CMOVL16rr: return X86::CMOVG16rr;
2967 case X86::CMOVL32rm: return X86::CMOVG32rm;
2968 case X86::CMOVL32rr: return X86::CMOVG32rr;
2969 case X86::CMOVL64rm: return X86::CMOVG64rm;
2970 case X86::CMOVL64rr: return X86::CMOVG64rr;
2971 case X86::CMOVLE16rm: return X86::CMOVGE16rm;
2972 case X86::CMOVLE16rr: return X86::CMOVGE16rr;
2973 case X86::CMOVLE32rm: return X86::CMOVGE32rm;
2974 case X86::CMOVLE32rr: return X86::CMOVGE32rr;
2975 case X86::CMOVLE64rm: return X86::CMOVGE64rm;
2976 case X86::CMOVLE64rr: return X86::CMOVGE64rr;
2977
2978 case X86::CMOVG16rm: return X86::CMOVL16rm;
2979 case X86::CMOVG16rr: return X86::CMOVL16rr;
2980 case X86::CMOVG32rm: return X86::CMOVL32rm;
2981 case X86::CMOVG32rr: return X86::CMOVL32rr;
2982 case X86::CMOVG64rm: return X86::CMOVL64rm;
2983 case X86::CMOVG64rr: return X86::CMOVL64rr;
2984 case X86::CMOVGE16rm: return X86::CMOVLE16rm;
2985 case X86::CMOVGE16rr: return X86::CMOVLE16rr;
2986 case X86::CMOVGE32rm: return X86::CMOVLE32rm;
2987 case X86::CMOVGE32rr: return X86::CMOVLE32rr;
2988 case X86::CMOVGE64rm: return X86::CMOVLE64rm;
2989 case X86::CMOVGE64rr: return X86::CMOVLE64rr;
2990
2991 case X86::CMOVB16rm: return X86::CMOVA16rm;
2992 case X86::CMOVB16rr: return X86::CMOVA16rr;
2993 case X86::CMOVB32rm: return X86::CMOVA32rm;
2994 case X86::CMOVB32rr: return X86::CMOVA32rr;
2995 case X86::CMOVB64rm: return X86::CMOVA64rm;
2996 case X86::CMOVB64rr: return X86::CMOVA64rr;
2997 case X86::CMOVBE16rm: return X86::CMOVAE16rm;
2998 case X86::CMOVBE16rr: return X86::CMOVAE16rr;
2999 case X86::CMOVBE32rm: return X86::CMOVAE32rm;
3000 case X86::CMOVBE32rr: return X86::CMOVAE32rr;
3001 case X86::CMOVBE64rm: return X86::CMOVAE64rm;
3002 case X86::CMOVBE64rr: return X86::CMOVAE64rr;
3003
3004 case X86::CMOVA16rm: return X86::CMOVB16rm;
3005 case X86::CMOVA16rr: return X86::CMOVB16rr;
3006 case X86::CMOVA32rm: return X86::CMOVB32rm;
3007 case X86::CMOVA32rr: return X86::CMOVB32rr;
3008 case X86::CMOVA64rm: return X86::CMOVB64rm;
3009 case X86::CMOVA64rr: return X86::CMOVB64rr;
3010 case X86::CMOVAE16rm: return X86::CMOVBE16rm;
3011 case X86::CMOVAE16rr: return X86::CMOVBE16rr;
3012 case X86::CMOVAE32rm: return X86::CMOVBE32rm;
3013 case X86::CMOVAE32rr: return X86::CMOVBE32rr;
3014 case X86::CMOVAE64rm: return X86::CMOVBE64rm;
3015 case X86::CMOVAE64rr: return X86::CMOVBE64rr;
3016 }
3017}
3018
3019/// isRedundantFlagInstr - check whether the first instruction, whose only
3020/// purpose is to update flags, can be made redundant.
3021/// CMPrr can be made redundant by SUBrr if the operands are the same.
3022/// This function can be extended later on.
3023/// SrcReg, SrcRegs: register operands for FlagI.
3024/// ImmValue: immediate for FlagI if it takes an immediate.
3025inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3026 unsigned SrcReg2, int ImmValue,
3027 MachineInstr *OI) {
3028 if (((FlagI->getOpcode() == X86::CMP64rr &&
3029 OI->getOpcode() == X86::SUB64rr) ||
3030 (FlagI->getOpcode() == X86::CMP32rr &&
3031 OI->getOpcode() == X86::SUB32rr)||
3032 (FlagI->getOpcode() == X86::CMP16rr &&
3033 OI->getOpcode() == X86::SUB16rr)||
3034 (FlagI->getOpcode() == X86::CMP8rr &&
3035 OI->getOpcode() == X86::SUB8rr)) &&
3036 ((OI->getOperand(1).getReg() == SrcReg &&
3037 OI->getOperand(2).getReg() == SrcReg2) ||
3038 (OI->getOperand(1).getReg() == SrcReg2 &&
3039 OI->getOperand(2).getReg() == SrcReg)))
3040 return true;
3041
3042 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3043 OI->getOpcode() == X86::SUB64ri32) ||
3044 (FlagI->getOpcode() == X86::CMP64ri8 &&
3045 OI->getOpcode() == X86::SUB64ri8) ||
3046 (FlagI->getOpcode() == X86::CMP32ri &&
3047 OI->getOpcode() == X86::SUB32ri) ||
3048 (FlagI->getOpcode() == X86::CMP32ri8 &&
3049 OI->getOpcode() == X86::SUB32ri8) ||
3050 (FlagI->getOpcode() == X86::CMP16ri &&
3051 OI->getOpcode() == X86::SUB16ri) ||
3052 (FlagI->getOpcode() == X86::CMP16ri8 &&
3053 OI->getOpcode() == X86::SUB16ri8) ||
3054 (FlagI->getOpcode() == X86::CMP8ri &&
3055 OI->getOpcode() == X86::SUB8ri)) &&
3056 OI->getOperand(1).getReg() == SrcReg &&
3057 OI->getOperand(2).getImm() == ImmValue)
3058 return true;
3059 return false;
3060}
3061
3062/// optimizeCompareInstr - Check if there exists an earlier instruction that
3063/// operates on the same source operands and sets flags in the same way as
3064/// Compare; remove Compare if possible.
3065bool X86InstrInfo::
3066optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3067 int CmpMask, int CmpValue,
3068 const MachineRegisterInfo *MRI) const {
3069 // Get the unique definition of SrcReg.
3070 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3071 if (!MI) return false;
3072
3073 // CmpInstr is the first instruction of the BB.
3074 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3075
3076 // We are searching for an earlier instruction that can make CmpInstr
3077 // redundant and that instruction will be saved in Sub.
3078 MachineInstr *Sub = NULL;
3079 const TargetRegisterInfo *TRI = &getRegisterInfo();
3080
3081 // We iterate backward, starting from the instruction before CmpInstr and
3082 // stop when reaching the definition of a source register or done with the BB.
3083 // RI points to the instruction before CmpInstr.
3084 // If the definition is in this basic block, RE points to the definition;
3085 // otherwise, RE is the rend of the basic block.
3086 MachineBasicBlock::reverse_iterator
3087 RI = MachineBasicBlock::reverse_iterator(I),
3088 RE = CmpInstr->getParent() == MI->getParent() ?
3089 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3090 CmpInstr->getParent()->rend();
3091 for (; RI != RE; ++RI) {
3092 MachineInstr *Instr = &*RI;
3093 // Check whether CmpInstr can be made redundant by the current instruction.
3094 if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
3095 Sub = Instr;
3096 break;
3097 }
3098
3099 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
3100 Instr->readsRegister(X86::EFLAGS, TRI))
3101 // This instruction modifies or uses EFLAGS.
3102 // We can't remove CmpInstr.
3103 return false;
3104 }
3105
3106 // Return false if no candidates exist.
3107 if (!Sub)
3108 return false;
3109
3110 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3111 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3112 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3113 for (++I; I != E; ++I) {
3114 const MachineInstr &Instr = *I;
3115 if (Instr.modifiesRegister(X86::EFLAGS, TRI))
3116 // It is safe to remove CmpInstr if EFLAGS is updated again.
3117 break;
3118
3119 if (!Instr.readsRegister(X86::EFLAGS, TRI))
3120 continue;
3121
3122 // EFLAGS is used by this instruction.
3123 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3124 Sub->getOperand(2).getReg() == SrcReg) {
3125
3126 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3127 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3128 unsigned NewOpc = getSwappedConditionForSET(Instr.getOpcode());
3129 if (!NewOpc) NewOpc = getSwappedConditionForBranch(Instr.getOpcode());
3130 if (!NewOpc) NewOpc = getSwappedConditionForCMov(Instr.getOpcode());
3131 if (!NewOpc) return false;
3132
3133 // Push the MachineInstr to OpsToUpdate.
3134 // If it is safe to remove CmpInstr, the condition code of these
3135 // instructions will be modified.
3136 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3137 }
3138 }
3139
3140 // Make sure Sub instruction defines EFLAGS.
3141 assert(Sub->getNumOperands() >= 4 && Sub->getOperand(3).isReg() &&
3142 Sub->getOperand(3).getReg() == X86::EFLAGS &&
3143 "EFLAGS should be the 4th operand of SUBrr or SUBri.");
3144 Sub->getOperand(3).setIsDef(true);
3145 CmpInstr->eraseFromParent();
3146
3147 // Modify the condition code of instructions in OpsToUpdate.
3148 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3149 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3150 return true;
3151}
3152
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3153/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3154/// instruction with two undef reads of the register being defined. This is
3155/// used for mapping:
3156/// %xmm4 = V_SET0
3157/// to:
3158/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3159///
3160static bool Expand2AddrUndef(MachineInstr *MI, const MCInstrDesc &Desc) {
3161 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3162 unsigned Reg = MI->getOperand(0).getReg();
3163 MI->setDesc(Desc);
3164
3165 // MachineInstr::addOperand() will insert explicit operands before any
3166 // implicit operands.
3167 MachineInstrBuilder(MI).addReg(Reg, RegState::Undef)
3168 .addReg(Reg, RegState::Undef);
3169 // But we don't trust that.
3170 assert(MI->getOperand(1).getReg() == Reg &&
3171 MI->getOperand(2).getReg() == Reg && "Misplaced operand");
3172 return true;
3173}
3174
3175bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3176 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3177 switch (MI->getOpcode()) {
3178 case X86::V_SET0:
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +0000[diff] [blame]3179 case X86::FsFLD0SS:
3180 case X86::FsFLD0SD:
Jakob Stoklund Olesen3e5d5c52011-11-07 19:15:58 +0000[diff] [blame]3181 return Expand2AddrUndef(MI, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
Jakob Stoklund Olesened744822011-10-08 18:28:28 +0000[diff] [blame]3182 case X86::TEST8ri_NOREX:
3183 MI->setDesc(get(X86::TEST8ri));
3184 return true;
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3185 }
3186 return false;
3187}
3188
Evan Cheng962021b2010-04-26 07:38:55 +0000[diff] [blame]3189MachineInstr*
3190X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
Evan Cheng8601a3d2010-04-29 01:13:30 +0000[diff] [blame]3191 int FrameIx, uint64_t Offset,
Evan Cheng962021b2010-04-26 07:38:55 +0000[diff] [blame]3192 const MDNode *MDPtr,
3193 DebugLoc DL) const {
Evan Cheng962021b2010-04-26 07:38:55 +0000[diff] [blame]3194 X86AddressMode AM;
3195 AM.BaseType = X86AddressMode::FrameIndexBase;
3196 AM.Base.FrameIndex = FrameIx;
3197 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
3198 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
3199 return &*MIB;
3200}
3201
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]3202static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
Dan Gohmand68a0762009-01-05 17:59:02 +0000[diff] [blame]3203 const SmallVectorImpl<MachineOperand> &MOs,
Bill Wendling9bc96a52009-02-03 00:55:04 +0000[diff] [blame]3204 MachineInstr *MI,
3205 const TargetInstrInfo &TII) {
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3206 // Create the base instruction with the memory operand as the first part.
Bill Wendling9bc96a52009-02-03 00:55:04 +0000[diff] [blame]3207 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3208 MI->getDebugLoc(), true);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3209 MachineInstrBuilder MIB(NewMI);
3210 unsigned NumAddrOps = MOs.size();
3211 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3212 MIB.addOperand(MOs[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3213 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]3214 addOffset(MIB, 0);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3215
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3216 // Loop over the rest of the ri operands, converting them over.
Chris Lattner749c6f62008-01-07 07:27:27 +0000[diff] [blame]3217 unsigned NumOps = MI->getDesc().getNumOperands()-2;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3218 for (unsigned i = 0; i != NumOps; ++i) {
3219 MachineOperand &MO = MI->getOperand(i+2);
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3220 MIB.addOperand(MO);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3221 }
3222 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3223 MachineOperand &MO = MI->getOperand(i);
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3224 MIB.addOperand(MO);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3225 }
3226 return MIB;
3227}
3228
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]3229static MachineInstr *FuseInst(MachineFunction &MF,
3230 unsigned Opcode, unsigned OpNo,
Dan Gohmand68a0762009-01-05 17:59:02 +0000[diff] [blame]3231 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3232 MachineInstr *MI, const TargetInstrInfo &TII) {
Bill Wendling9bc96a52009-02-03 00:55:04 +0000[diff] [blame]3233 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3234 MI->getDebugLoc(), true);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3235 MachineInstrBuilder MIB(NewMI);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3236
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3237 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3238 MachineOperand &MO = MI->getOperand(i);
3239 if (i == OpNo) {
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]3240 assert(MO.isReg() && "Expected to fold into reg operand!");
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3241 unsigned NumAddrOps = MOs.size();
3242 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3243 MIB.addOperand(MOs[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3244 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]3245 addOffset(MIB, 0);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3246 } else {
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3247 MIB.addOperand(MO);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3248 }
3249 }
3250 return MIB;
3251}
3252
3253static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
Dan Gohmand68a0762009-01-05 17:59:02 +0000[diff] [blame]3254 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3255 MachineInstr *MI) {
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]3256 MachineFunction &MF = *MI->getParent()->getParent();
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]3257 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3258
3259 unsigned NumAddrOps = MOs.size();
3260 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3261 MIB.addOperand(MOs[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3262 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]3263 addOffset(MIB, 0);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3264 return MIB.addImm(0);
3265}
3266
3267MachineInstr*
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3268X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3269 MachineInstr *MI, unsigned i,
Evan Chengf9b36f02009-07-15 06:10:07 +0000[diff] [blame]3270 const SmallVectorImpl<MachineOperand> &MOs,
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3271 unsigned Size, unsigned Align) const {
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3272 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3273 bool isTwoAddrFold = false;
Chris Lattner749c6f62008-01-07 07:27:27 +0000[diff] [blame]3274 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3275 bool isTwoAddr = NumOps > 1 &&
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3276 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3277
Jakob Stoklund Olesen60045c22011-04-30 23:00:05 +0000[diff] [blame]3278 // FIXME: AsmPrinter doesn't know how to handle
3279 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3280 if (MI->getOpcode() == X86::ADD32ri &&
3281 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3282 return NULL;
3283
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3284 MachineInstr *NewMI = NULL;
3285 // Folding a memory location into the two-address part of a two-address
3286 // instruction is different than folding it other places. It requires
3287 // replacing the *two* registers with the memory location.
3288 if (isTwoAddr && NumOps >= 2 && i < 2 &&
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]3289 MI->getOperand(0).isReg() &&
3290 MI->getOperand(1).isReg() &&
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3291 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3292 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3293 isTwoAddrFold = true;
3294 } else if (i == 0) { // If operand 0
Dan Gohmanf1b4d262010-01-12 04:42:54 +0000[diff] [blame]3295 if (MI->getOpcode() == X86::MOV64r0)
3296 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
3297 else if (MI->getOpcode() == X86::MOV32r0)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3298 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
Dan Gohmanf1b4d262010-01-12 04:42:54 +0000[diff] [blame]3299 else if (MI->getOpcode() == X86::MOV16r0)
3300 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3301 else if (MI->getOpcode() == X86::MOV8r0)
3302 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]3303 if (NewMI)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3304 return NewMI;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3305
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3306 OpcodeTablePtr = &RegOp2MemOpTable0;
3307 } else if (i == 1) {
3308 OpcodeTablePtr = &RegOp2MemOpTable1;
3309 } else if (i == 2) {
3310 OpcodeTablePtr = &RegOp2MemOpTable2;
3311 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3312
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3313 // If table selected...
3314 if (OpcodeTablePtr) {
3315 // Find the Opcode to fuse
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3316 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3317 OpcodeTablePtr->find(MI->getOpcode());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3318 if (I != OpcodeTablePtr->end()) {
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3319 unsigned Opcode = I->second.first;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]3320 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
Evan Chengf9b36f02009-07-15 06:10:07 +0000[diff] [blame]3321 if (Align < MinAlign)
3322 return NULL;
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3323 bool NarrowToMOV32rm = false;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3324 if (Size) {
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]3325 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3326 if (Size < RCSize) {
3327 // Check if it's safe to fold the load. If the size of the object is
3328 // narrower than the load width, then it's not.
3329 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3330 return NULL;
3331 // If this is a 64-bit load, but the spill slot is 32, then we can do
3332 // a 32-bit load which is implicitly zero-extended. This likely is due
3333 // to liveintervalanalysis remat'ing a load from stack slot.
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3334 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3335 return NULL;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3336 Opcode = X86::MOV32rm;
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3337 NarrowToMOV32rm = true;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3338 }
3339 }
3340
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3341 if (isTwoAddrFold)
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3342 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3343 else
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3344 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3345
3346 if (NarrowToMOV32rm) {
3347 // If this is the special case where we use a MOV32rm to load a 32-bit
3348 // value and zero-extend the top bits. Change the destination register
3349 // to a 32-bit one.
3350 unsigned DstReg = NewMI->getOperand(0).getReg();
3351 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3352 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
Jakob Stoklund Olesen3458e9e2010-05-24 14:48:17 +0000[diff] [blame]3353 X86::sub_32bit));
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3354 else
Jakob Stoklund Olesen3458e9e2010-05-24 14:48:17 +0000[diff] [blame]3355 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3356 }
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3357 return NewMI;
3358 }
3359 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3360
3361 // No fusion
Jakob Stoklund Olesen9c50e8b2010-07-09 20:43:09 +0000[diff] [blame]3362 if (PrintFailedFusing && !MI->isCopy())
David Greene5b901322010-01-05 01:29:29 +0000[diff] [blame]3363 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3364 return NULL;
3365}
3366
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3367/// hasPartialRegUpdate - Return true for all instructions that only update
3368/// the first 32 or 64-bits of the destination register and leave the rest
3369/// unmodified. This can be used to avoid folding loads if the instructions
3370/// only update part of the destination register, and the non-updated part is
3371/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3372/// instructions breaks the partial register dependency and it can improve
3373/// performance. e.g.:
3374///
3375/// movss (%rdi), %xmm0
3376/// cvtss2sd %xmm0, %xmm0
3377///
3378/// Instead of
3379/// cvtss2sd (%rdi), %xmm0
3380///
Bruno Cardoso Lopes6b5b79c2011-09-15 23:04:24 +0000[diff] [blame]3381/// FIXME: This should be turned into a TSFlags.
3382///
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3383static bool hasPartialRegUpdate(unsigned Opcode) {
3384 switch (Opcode) {
Jakob Stoklund Olesenc2ecf3e2011-11-15 01:15:30 +0000[diff] [blame]3385 case X86::CVTSI2SSrr:
3386 case X86::CVTSI2SS64rr:
3387 case X86::CVTSI2SDrr:
3388 case X86::CVTSI2SD64rr:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3389 case X86::CVTSD2SSrr:
3390 case X86::Int_CVTSD2SSrr:
3391 case X86::CVTSS2SDrr:
3392 case X86::Int_CVTSS2SDrr:
3393 case X86::RCPSSr:
3394 case X86::RCPSSr_Int:
3395 case X86::ROUNDSDr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +0000[diff] [blame]3396 case X86::ROUNDSDr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3397 case X86::ROUNDSSr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +0000[diff] [blame]3398 case X86::ROUNDSSr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3399 case X86::RSQRTSSr:
3400 case X86::RSQRTSSr_Int:
3401 case X86::SQRTSSr:
3402 case X86::SQRTSSr_Int:
3403 // AVX encoded versions
3404 case X86::VCVTSD2SSrr:
3405 case X86::Int_VCVTSD2SSrr:
3406 case X86::VCVTSS2SDrr:
3407 case X86::Int_VCVTSS2SDrr:
3408 case X86::VRCPSSr:
3409 case X86::VROUNDSDr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +0000[diff] [blame]3410 case X86::VROUNDSDr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3411 case X86::VROUNDSSr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +0000[diff] [blame]3412 case X86::VROUNDSSr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3413 case X86::VRSQRTSSr:
3414 case X86::VSQRTSSr:
3415 return true;
3416 }
3417
3418 return false;
3419}
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3420
Jakob Stoklund Olesenc2ecf3e2011-11-15 01:15:30 +0000[diff] [blame]3421/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3422/// instructions we would like before a partial register update.
3423unsigned X86InstrInfo::
3424getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3425 const TargetRegisterInfo *TRI) const {
3426 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3427 return 0;
3428
3429 // If MI is marked as reading Reg, the partial register update is wanted.
3430 const MachineOperand &MO = MI->getOperand(0);
3431 unsigned Reg = MO.getReg();
3432 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3433 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3434 return 0;
3435 } else {
3436 if (MI->readsRegister(Reg, TRI))
3437 return 0;
3438 }
3439
3440 // If any of the preceding 16 instructions are reading Reg, insert a
3441 // dependency breaking instruction. The magic number is based on a few
3442 // Nehalem experiments.
3443 return 16;
3444}
3445
3446void X86InstrInfo::
3447breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3448 const TargetRegisterInfo *TRI) const {
3449 unsigned Reg = MI->getOperand(OpNum).getReg();
3450 if (X86::VR128RegClass.contains(Reg)) {
3451 // These instructions are all floating point domain, so xorps is the best
3452 // choice.
3453 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3454 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3455 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3456 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3457 } else if (X86::VR256RegClass.contains(Reg)) {
3458 // Use vxorps to clear the full ymm register.
3459 // It wants to read and write the xmm sub-register.
3460 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3461 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3462 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3463 .addReg(Reg, RegState::ImplicitDefine);
3464 } else
3465 return;
3466 MI->addRegisterKilled(Reg, TRI, true);
3467}
3468
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3469MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3470 MachineInstr *MI,
Evan Chengf9b36f02009-07-15 06:10:07 +0000[diff] [blame]3471 const SmallVectorImpl<unsigned> &Ops,
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3472 int FrameIndex) const {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3473 // Check switch flag
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3474 if (NoFusing) return NULL;
3475
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3476 // Unless optimizing for size, don't fold to avoid partial
3477 // register update stalls
3478 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3479 hasPartialRegUpdate(MI->getOpcode()))
3480 return 0;
Evan Cheng400073d2009-12-18 07:40:29 +0000[diff] [blame]3481
Evan Cheng5fd79d02008-02-08 21:20:40 +0000[diff] [blame]3482 const MachineFrameInfo *MFI = MF.getFrameInfo();
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3483 unsigned Size = MFI->getObjectSize(FrameIndex);
Evan Cheng5fd79d02008-02-08 21:20:40 +0000[diff] [blame]3484 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3485 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3486 unsigned NewOpc = 0;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3487 unsigned RCSize = 0;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3488 switch (MI->getOpcode()) {
3489 default: return NULL;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3490 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
Dan Gohmane5efbaf2010-05-18 21:42:03 +0000[diff] [blame]3491 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
3492 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
3493 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3494 }
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3495 // Check if it's safe to fold the load. If the size of the object is
3496 // narrower than the load width, then it's not.
3497 if (Size < RCSize)
3498 return NULL;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3499 // Change to CMPXXri r, 0 first.
Chris Lattner5080f4d2008-01-11 18:10:50 +0000[diff] [blame]3500 MI->setDesc(get(NewOpc));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3501 MI->getOperand(1).ChangeToImmediate(0);
3502 } else if (Ops.size() != 1)
3503 return NULL;
3504
3505 SmallVector<MachineOperand,4> MOs;
3506 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3507 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3508}
3509
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3510MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3511 MachineInstr *MI,
Evan Chengf9b36f02009-07-15 06:10:07 +0000[diff] [blame]3512 const SmallVectorImpl<unsigned> &Ops,
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3513 MachineInstr *LoadMI) const {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3514 // Check switch flag
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3515 if (NoFusing) return NULL;
3516
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3517 // Unless optimizing for size, don't fold to avoid partial
3518 // register update stalls
3519 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3520 hasPartialRegUpdate(MI->getOpcode()))
3521 return 0;
Evan Cheng400073d2009-12-18 07:40:29 +0000[diff] [blame]3522
Dan Gohmancddc11e2008-07-12 00:10:52 +0000[diff] [blame]3523 // Determine the alignment of the load.
Evan Cheng5fd79d02008-02-08 21:20:40 +0000[diff] [blame]3524 unsigned Alignment = 0;
Dan Gohmancddc11e2008-07-12 00:10:52 +0000[diff] [blame]3525 if (LoadMI->hasOneMemOperand())
Dan Gohmanc76909a2009-09-25 20:36:54 +0000[diff] [blame]3526 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3527 else
3528 switch (LoadMI->getOpcode()) {
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]3529 case X86::AVX_SET0PSY:
3530 case X86::AVX_SET0PDY:
Craig Topper745a86b2011-11-19 22:34:59 +0000[diff] [blame]3531 case X86::AVX2_SETALLONES:
Craig Topper12216172012-01-13 08:12:35 +0000[diff] [blame]3532 case X86::AVX2_SET0:
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]3533 Alignment = 32;
3534 break;
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3535 case X86::V_SET0:
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3536 case X86::V_SETALLONES:
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +0000[diff] [blame]3537 case X86::AVX_SETALLONES:
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3538 Alignment = 16;
3539 break;
3540 case X86::FsFLD0SD:
3541 Alignment = 8;
3542 break;
3543 case X86::FsFLD0SS:
3544 Alignment = 4;
3545 break;
3546 default:
Eli Friedmanbe5cbaa2011-06-10 01:13:01 +0000[diff] [blame]3547 return 0;
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3548 }
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3549 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3550 unsigned NewOpc = 0;
3551 switch (MI->getOpcode()) {
3552 default: return NULL;
3553 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]3554 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
3555 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
3556 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3557 }
3558 // Change to CMPXXri r, 0 first.
Chris Lattner5080f4d2008-01-11 18:10:50 +0000[diff] [blame]3559 MI->setDesc(get(NewOpc));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3560 MI->getOperand(1).ChangeToImmediate(0);
3561 } else if (Ops.size() != 1)
3562 return NULL;
3563
Jakob Stoklund Olesend29583b2010-08-11 23:08:22 +0000[diff] [blame]3564 // Make sure the subregisters match.
3565 // Otherwise we risk changing the size of the load.
3566 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
3567 return NULL;
3568
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]3569 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3570 switch (LoadMI->getOpcode()) {
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3571 case X86::V_SET0:
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3572 case X86::V_SETALLONES:
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]3573 case X86::AVX_SET0PSY:
3574 case X86::AVX_SET0PDY:
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +0000[diff] [blame]3575 case X86::AVX_SETALLONES:
Craig Topper745a86b2011-11-19 22:34:59 +0000[diff] [blame]3576 case X86::AVX2_SETALLONES:
Craig Topper12216172012-01-13 08:12:35 +0000[diff] [blame]3577 case X86::AVX2_SET0:
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3578 case X86::FsFLD0SD:
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +0000[diff] [blame]3579 case X86::FsFLD0SS: {
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3580 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3581 // Create a constant-pool entry and operands to load from it.
3582
Dan Gohman81d0c362010-03-09 03:01:40 +0000[diff] [blame]3583 // Medium and large mode can't fold loads this way.
3584 if (TM.getCodeModel() != CodeModel::Small &&
3585 TM.getCodeModel() != CodeModel::Kernel)
3586 return NULL;
3587
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3588 // x86-32 PIC requires a PIC base register for constant pools.
3589 unsigned PICBase = 0;
Jakob Stoklund Olesen93e55de2009-07-16 21:24:13 +0000[diff] [blame]3590 if (TM.getRelocationModel() == Reloc::PIC_) {
Evan Cheng2b48ab92009-07-16 18:44:05 +0000[diff] [blame]3591 if (TM.getSubtarget<X86Subtarget>().is64Bit())
3592 PICBase = X86::RIP;
Jakob Stoklund Olesen93e55de2009-07-16 21:24:13 +0000[diff] [blame]3593 else
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]3594 // FIXME: PICBase = getGlobalBaseReg(&MF);
Evan Cheng2b48ab92009-07-16 18:44:05 +0000[diff] [blame]3595 // This doesn't work for several reasons.
3596 // 1. GlobalBaseReg may have been spilled.
3597 // 2. It may not be live at MI.
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3598 return NULL;
Jakob Stoklund Olesen93e55de2009-07-16 21:24:13 +0000[diff] [blame]3599 }
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3600
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3601 // Create a constant-pool entry.
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3602 MachineConstantPool &MCP = *MF.getConstantPool();
Chris Lattnerdb125cf2011-07-18 04:54:35 +0000[diff] [blame]3603 Type *Ty;
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]3604 unsigned Opc = LoadMI->getOpcode();
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +0000[diff] [blame]3605 if (Opc == X86::FsFLD0SS)
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3606 Ty = Type::getFloatTy(MF.getFunction()->getContext());
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +0000[diff] [blame]3607 else if (Opc == X86::FsFLD0SD)
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3608 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]3609 else if (Opc == X86::AVX_SET0PSY || Opc == X86::AVX_SET0PDY)
3610 Ty = VectorType::get(Type::getFloatTy(MF.getFunction()->getContext()), 8);
Craig Topper12216172012-01-13 08:12:35 +0000[diff] [blame]3611 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX2_SET0)
Craig Topperb9c7f652012-01-13 06:12:41 +0000[diff] [blame]3612 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3613 else
3614 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +0000[diff] [blame]3615
Craig Topper745a86b2011-11-19 22:34:59 +0000[diff] [blame]3616 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX_SETALLONES ||
3617 Opc == X86::AVX2_SETALLONES);
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +0000[diff] [blame]3618 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
3619 Constant::getNullValue(Ty);
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3620 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3621
3622 // Create operands to load from the constant pool entry.
3623 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
3624 MOs.push_back(MachineOperand::CreateImm(1));
3625 MOs.push_back(MachineOperand::CreateReg(0, false));
3626 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]3627 MOs.push_back(MachineOperand::CreateReg(0, false));
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3628 break;
3629 }
3630 default: {
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3631 // Folding a normal load. Just copy the load's address operands.
3632 unsigned NumOps = LoadMI->getDesc().getNumOperands();
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]3633 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3634 MOs.push_back(LoadMI->getOperand(i));
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3635 break;
3636 }
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3637 }
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3638 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3639}
3640
3641
Dan Gohman8e8b8a22008-10-16 01:49:15 +0000[diff] [blame]3642bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
3643 const SmallVectorImpl<unsigned> &Ops) const {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3644 // Check switch flag
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3645 if (NoFusing) return 0;
3646
3647 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3648 switch (MI->getOpcode()) {
3649 default: return false;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3650 case X86::TEST8rr:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3651 case X86::TEST16rr:
3652 case X86::TEST32rr:
3653 case X86::TEST64rr:
3654 return true;
Jakob Stoklund Olesen60045c22011-04-30 23:00:05 +0000[diff] [blame]3655 case X86::ADD32ri:
3656 // FIXME: AsmPrinter doesn't know how to handle
3657 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3658 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3659 return false;
3660 break;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3661 }
3662 }
3663
3664 if (Ops.size() != 1)
3665 return false;
3666
3667 unsigned OpNum = Ops[0];
3668 unsigned Opc = MI->getOpcode();
Chris Lattner749c6f62008-01-07 07:27:27 +0000[diff] [blame]3669 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3670 bool isTwoAddr = NumOps > 1 &&
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3671 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3672
3673 // Folding a memory location into the two-address part of a two-address
3674 // instruction is different than folding it other places. It requires
3675 // replacing the *two* registers with the memory location.
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3676 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3677 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3678 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3679 } else if (OpNum == 0) { // If operand 0
3680 switch (Opc) {
Chris Lattner9ac75422009-07-14 20:19:57 +0000[diff] [blame]3681 case X86::MOV8r0:
Dan Gohmanf1b4d262010-01-12 04:42:54 +0000[diff] [blame]3682 case X86::MOV16r0:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3683 case X86::MOV32r0:
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3684 case X86::MOV64r0: return true;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3685 default: break;
3686 }
3687 OpcodeTablePtr = &RegOp2MemOpTable0;
3688 } else if (OpNum == 1) {
3689 OpcodeTablePtr = &RegOp2MemOpTable1;
3690 } else if (OpNum == 2) {
3691 OpcodeTablePtr = &RegOp2MemOpTable2;
3692 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3693
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]3694 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
3695 return true;
Jakob Stoklund Olesen1f323402010-07-09 20:43:13 +0000[diff] [blame]3696 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3697}
3698
3699bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
3700 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]3701 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3702 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3703 MemOp2RegOpTable.find(MI->getOpcode());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3704 if (I == MemOp2RegOpTable.end())
3705 return false;
3706 unsigned Opc = I->second.first;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]3707 unsigned Index = I->second.second & TB_INDEX_MASK;
3708 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3709 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3710 if (UnfoldLoad && !FoldedLoad)
3711 return false;
3712 UnfoldLoad &= FoldedLoad;
3713 if (UnfoldStore && !FoldedStore)
3714 return false;
3715 UnfoldStore &= FoldedStore;
3716
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3717 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]3718 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng98ec91e2010-07-02 20:36:18 +0000[diff] [blame]3719 if (!MI->hasOneMemOperand() &&
3720 RC == &X86::VR128RegClass &&
3721 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3722 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
3723 // conservatively assume the address is unaligned. That's bad for
3724 // performance.
3725 return false;
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]3726 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3727 SmallVector<MachineOperand,2> BeforeOps;
3728 SmallVector<MachineOperand,2> AfterOps;
3729 SmallVector<MachineOperand,4> ImpOps;
3730 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3731 MachineOperand &Op = MI->getOperand(i);
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]3732 if (i >= Index && i < Index + X86::AddrNumOperands)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3733 AddrOps.push_back(Op);
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]3734 else if (Op.isReg() && Op.isImplicit())
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3735 ImpOps.push_back(Op);
3736 else if (i < Index)
3737 BeforeOps.push_back(Op);
3738 else if (i > Index)
3739 AfterOps.push_back(Op);
3740 }
3741
3742 // Emit the load instruction.
3743 if (UnfoldLoad) {
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]3744 std::pair<MachineInstr::mmo_iterator,
3745 MachineInstr::mmo_iterator> MMOs =
3746 MF.extractLoadMemRefs(MI->memoperands_begin(),
3747 MI->memoperands_end());
3748 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3749 if (UnfoldStore) {
3750 // Address operands cannot be marked isKill.
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]3751 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3752 MachineOperand &MO = NewMIs[0]->getOperand(i);
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]3753 if (MO.isReg())
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3754 MO.setIsKill(false);
3755 }
3756 }
3757 }
3758
3759 // Emit the data processing instruction.
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3760 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3761 MachineInstrBuilder MIB(DataMI);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3762
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3763 if (FoldedStore)
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]3764 MIB.addReg(Reg, RegState::Define);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3765 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3766 MIB.addOperand(BeforeOps[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3767 if (FoldedLoad)
3768 MIB.addReg(Reg);
3769 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3770 MIB.addOperand(AfterOps[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3771 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
3772 MachineOperand &MO = ImpOps[i];
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]3773 MIB.addReg(MO.getReg(),
3774 getDefRegState(MO.isDef()) |
3775 RegState::Implicit |
3776 getKillRegState(MO.isKill()) |
Evan Cheng4784f1f2009-06-30 08:49:04 +0000[diff] [blame]3777 getDeadRegState(MO.isDead()) |
3778 getUndefRegState(MO.isUndef()));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3779 }
3780 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
3781 unsigned NewOpc = 0;
3782 switch (DataMI->getOpcode()) {
3783 default: break;
3784 case X86::CMP64ri32:
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]3785 case X86::CMP64ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3786 case X86::CMP32ri:
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]3787 case X86::CMP32ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3788 case X86::CMP16ri:
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]3789 case X86::CMP16ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3790 case X86::CMP8ri: {
3791 MachineOperand &MO0 = DataMI->getOperand(0);
3792 MachineOperand &MO1 = DataMI->getOperand(1);
3793 if (MO1.getImm() == 0) {
3794 switch (DataMI->getOpcode()) {
3795 default: break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]3796 case X86::CMP64ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3797 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]3798 case X86::CMP32ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3799 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]3800 case X86::CMP16ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3801 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
3802 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
3803 }
Chris Lattner5080f4d2008-01-11 18:10:50 +0000[diff] [blame]3804 DataMI->setDesc(get(NewOpc));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3805 MO1.ChangeToRegister(MO0.getReg(), false);
3806 }
3807 }
3808 }
3809 NewMIs.push_back(DataMI);
3810
3811 // Emit the store instruction.
3812 if (UnfoldStore) {
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]3813 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]3814 std::pair<MachineInstr::mmo_iterator,
3815 MachineInstr::mmo_iterator> MMOs =
3816 MF.extractStoreMemRefs(MI->memoperands_begin(),
3817 MI->memoperands_end());
3818 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3819 }
3820
3821 return true;
3822}
3823
3824bool
3825X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]3826 SmallVectorImpl<SDNode*> &NewNodes) const {
Dan Gohmane8be6c62008-07-17 19:10:17 +0000[diff] [blame]3827 if (!N->isMachineOpcode())
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3828 return false;
3829
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3830 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3831 MemOp2RegOpTable.find(N->getMachineOpcode());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3832 if (I == MemOp2RegOpTable.end())
3833 return false;
3834 unsigned Opc = I->second.first;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]3835 unsigned Index = I->second.second & TB_INDEX_MASK;
3836 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3837 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3838 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]3839 MachineFunction &MF = DAG.getMachineFunction();
3840 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3841 unsigned NumDefs = MCID.NumDefs;
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]3842 std::vector<SDValue> AddrOps;
3843 std::vector<SDValue> BeforeOps;
3844 std::vector<SDValue> AfterOps;
Dale Johannesened2eee62009-02-06 01:31:28 +0000[diff] [blame]3845 DebugLoc dl = N->getDebugLoc();
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3846 unsigned NumOps = N->getNumOperands();
Dan Gohmanc76909a2009-09-25 20:36:54 +0000[diff] [blame]3847 for (unsigned i = 0; i != NumOps-1; ++i) {
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]3848 SDValue Op = N->getOperand(i);
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]3849 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3850 AddrOps.push_back(Op);
Dan Gohmanb37a8202009-03-04 19:23:38 +0000[diff] [blame]3851 else if (i < Index-NumDefs)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3852 BeforeOps.push_back(Op);
Dan Gohmanb37a8202009-03-04 19:23:38 +0000[diff] [blame]3853 else if (i > Index-NumDefs)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3854 AfterOps.push_back(Op);
3855 }
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]3856 SDValue Chain = N->getOperand(NumOps-1);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3857 AddrOps.push_back(Chain);
3858
3859 // Emit the load instruction.
3860 SDNode *Load = 0;
3861 if (FoldedLoad) {
Owen Andersone50ed302009-08-10 22:56:29 +0000[diff] [blame]3862 EVT VT = *RC->vt_begin();
Evan Cheng600c0432009-11-16 21:56:03 +0000[diff] [blame]3863 std::pair<MachineInstr::mmo_iterator,
3864 MachineInstr::mmo_iterator> MMOs =
3865 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
3866 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng98ec91e2010-07-02 20:36:18 +0000[diff] [blame]3867 if (!(*MMOs.first) &&
3868 RC == &X86::VR128RegClass &&
3869 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3870 // Do not introduce a slow unaligned load.
3871 return false;
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]3872 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3873 bool isAligned = (*MMOs.first) &&
3874 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman602b0c82009-09-25 18:54:59 +0000[diff] [blame]3875 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
3876 VT, MVT::Other, &AddrOps[0], AddrOps.size());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3877 NewNodes.push_back(Load);
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]3878
3879 // Preserve memory reference information.
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]3880 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3881 }
3882
3883 // Emit the data processing instruction.
Owen Andersone50ed302009-08-10 22:56:29 +0000[diff] [blame]3884 std::vector<EVT> VTs;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3885 const TargetRegisterClass *DstRC = 0;
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3886 if (MCID.getNumDefs() > 0) {
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]3887 DstRC = getRegClass(MCID, 0, &RI, MF);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3888 VTs.push_back(*DstRC->vt_begin());
3889 }
3890 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
Owen Andersone50ed302009-08-10 22:56:29 +0000[diff] [blame]3891 EVT VT = N->getValueType(i);
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3892 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3893 VTs.push_back(VT);
3894 }
3895 if (Load)
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]3896 BeforeOps.push_back(SDValue(Load, 0));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3897 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
Dan Gohman602b0c82009-09-25 18:54:59 +0000[diff] [blame]3898 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0],
3899 BeforeOps.size());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3900 NewNodes.push_back(NewNode);
3901
3902 // Emit the store instruction.
3903 if (FoldedStore) {
3904 AddrOps.pop_back();
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]3905 AddrOps.push_back(SDValue(NewNode, 0));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3906 AddrOps.push_back(Chain);
Evan Cheng600c0432009-11-16 21:56:03 +0000[diff] [blame]3907 std::pair<MachineInstr::mmo_iterator,
3908 MachineInstr::mmo_iterator> MMOs =
3909 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
3910 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng98ec91e2010-07-02 20:36:18 +0000[diff] [blame]3911 if (!(*MMOs.first) &&
3912 RC == &X86::VR128RegClass &&
3913 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3914 // Do not introduce a slow unaligned store.
3915 return false;
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]3916 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3917 bool isAligned = (*MMOs.first) &&
3918 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman602b0c82009-09-25 18:54:59 +0000[diff] [blame]3919 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
3920 isAligned, TM),
3921 dl, MVT::Other,
3922 &AddrOps[0], AddrOps.size());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3923 NewNodes.push_back(Store);
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]3924
3925 // Preserve memory reference information.
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]3926 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3927 }
3928
3929 return true;
3930}
3931
3932unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
Dan Gohman0115e162009-10-30 22:18:41 +0000[diff] [blame]3933 bool UnfoldLoad, bool UnfoldStore,
3934 unsigned *LoadRegIndex) const {
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3935 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3936 MemOp2RegOpTable.find(Opc);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3937 if (I == MemOp2RegOpTable.end())
3938 return 0;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]3939 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3940 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3941 if (UnfoldLoad && !FoldedLoad)
3942 return 0;
3943 if (UnfoldStore && !FoldedStore)
3944 return 0;
Dan Gohman0115e162009-10-30 22:18:41 +0000[diff] [blame]3945 if (LoadRegIndex)
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]3946 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3947 return I->second.first;
3948}
3949
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]3950bool
3951X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
3952 int64_t &Offset1, int64_t &Offset2) const {
3953 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
3954 return false;
3955 unsigned Opc1 = Load1->getMachineOpcode();
3956 unsigned Opc2 = Load2->getMachineOpcode();
3957 switch (Opc1) {
3958 default: return false;
3959 case X86::MOV8rm:
3960 case X86::MOV16rm:
3961 case X86::MOV32rm:
3962 case X86::MOV64rm:
3963 case X86::LD_Fp32m:
3964 case X86::LD_Fp64m:
3965 case X86::LD_Fp80m:
3966 case X86::MOVSSrm:
3967 case X86::MOVSDrm:
3968 case X86::MMX_MOVD64rm:
3969 case X86::MMX_MOVQ64rm:
3970 case X86::FsMOVAPSrm:
3971 case X86::FsMOVAPDrm:
3972 case X86::MOVAPSrm:
3973 case X86::MOVUPSrm:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]3974 case X86::MOVAPDrm:
3975 case X86::MOVDQArm:
3976 case X86::MOVDQUrm:
Bruno Cardoso Lopesb4e905d2011-09-15 22:15:52 +0000[diff] [blame]3977 // AVX load instructions
3978 case X86::VMOVSSrm:
3979 case X86::VMOVSDrm:
3980 case X86::FsVMOVAPSrm:
3981 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]3982 case X86::VMOVAPSrm:
3983 case X86::VMOVUPSrm:
3984 case X86::VMOVAPDrm:
3985 case X86::VMOVDQArm:
3986 case X86::VMOVDQUrm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]3987 case X86::VMOVAPSYrm:
3988 case X86::VMOVUPSYrm:
3989 case X86::VMOVAPDYrm:
3990 case X86::VMOVDQAYrm:
3991 case X86::VMOVDQUYrm:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]3992 break;
3993 }
3994 switch (Opc2) {
3995 default: return false;
3996 case X86::MOV8rm:
3997 case X86::MOV16rm:
3998 case X86::MOV32rm:
3999 case X86::MOV64rm:
4000 case X86::LD_Fp32m:
4001 case X86::LD_Fp64m:
4002 case X86::LD_Fp80m:
4003 case X86::MOVSSrm:
4004 case X86::MOVSDrm:
4005 case X86::MMX_MOVD64rm:
4006 case X86::MMX_MOVQ64rm:
4007 case X86::FsMOVAPSrm:
4008 case X86::FsMOVAPDrm:
4009 case X86::MOVAPSrm:
4010 case X86::MOVUPSrm:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4011 case X86::MOVAPDrm:
4012 case X86::MOVDQArm:
4013 case X86::MOVDQUrm:
Bruno Cardoso Lopesb4e905d2011-09-15 22:15:52 +0000[diff] [blame]4014 // AVX load instructions
4015 case X86::VMOVSSrm:
4016 case X86::VMOVSDrm:
4017 case X86::FsVMOVAPSrm:
4018 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]4019 case X86::VMOVAPSrm:
4020 case X86::VMOVUPSrm:
4021 case X86::VMOVAPDrm:
4022 case X86::VMOVDQArm:
4023 case X86::VMOVDQUrm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]4024 case X86::VMOVAPSYrm:
4025 case X86::VMOVUPSYrm:
4026 case X86::VMOVAPDYrm:
4027 case X86::VMOVDQAYrm:
4028 case X86::VMOVDQUYrm:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4029 break;
4030 }
4031
4032 // Check if chain operands and base addresses match.
4033 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4034 Load1->getOperand(5) != Load2->getOperand(5))
4035 return false;
4036 // Segment operands should match as well.
4037 if (Load1->getOperand(4) != Load2->getOperand(4))
4038 return false;
4039 // Scale should be 1, Index should be Reg0.
4040 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4041 Load1->getOperand(2) == Load2->getOperand(2)) {
4042 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4043 return false;
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4044
4045 // Now let's examine the displacements.
4046 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4047 isa<ConstantSDNode>(Load2->getOperand(3))) {
4048 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4049 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4050 return true;
4051 }
4052 }
4053 return false;
4054}
4055
4056bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4057 int64_t Offset1, int64_t Offset2,
4058 unsigned NumLoads) const {
4059 assert(Offset2 > Offset1);
4060 if ((Offset2 - Offset1) / 8 > 64)
4061 return false;
4062
4063 unsigned Opc1 = Load1->getMachineOpcode();
4064 unsigned Opc2 = Load2->getMachineOpcode();
4065 if (Opc1 != Opc2)
4066 return false; // FIXME: overly conservative?
4067
4068 switch (Opc1) {
4069 default: break;
4070 case X86::LD_Fp32m:
4071 case X86::LD_Fp64m:
4072 case X86::LD_Fp80m:
4073 case X86::MMX_MOVD64rm:
4074 case X86::MMX_MOVQ64rm:
4075 return false;
4076 }
4077
4078 EVT VT = Load1->getValueType(0);
4079 switch (VT.getSimpleVT().SimpleTy) {
Bill Wendling19d85972010-06-22 22:16:17 +0000[diff] [blame]4080 default:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4081 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4082 // have 16 of them to play with.
4083 if (TM.getSubtargetImpl()->is64Bit()) {
4084 if (NumLoads >= 3)
4085 return false;
Bill Wendling19d85972010-06-22 22:16:17 +0000[diff] [blame]4086 } else if (NumLoads) {
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4087 return false;
Bill Wendling19d85972010-06-22 22:16:17 +0000[diff] [blame]4088 }
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4089 break;
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4090 case MVT::i8:
4091 case MVT::i16:
4092 case MVT::i32:
4093 case MVT::i64:
Evan Chengafc36732010-01-22 23:49:11 +0000[diff] [blame]4094 case MVT::f32:
4095 case MVT::f64:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4096 if (NumLoads)
4097 return false;
Bill Wendling19d85972010-06-22 22:16:17 +0000[diff] [blame]4098 break;
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4099 }
4100
4101 return true;
4102}
4103
4104
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]4105bool X86InstrInfo::
Owen Anderson44eb65c2008-08-14 22:49:33 +0000[diff] [blame]4106ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]4107 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
Evan Cheng97af60b2008-08-29 23:21:31 +0000[diff] [blame]4108 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]4109 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4110 return true;
Evan Cheng97af60b2008-08-29 23:21:31 +0000[diff] [blame]4111 Cond[0].setImm(GetOppositeBranchCondition(CC));
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]4112 return false;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]4113}
4114
Evan Cheng23066282008-10-27 07:14:50 +0000[diff] [blame]4115bool X86InstrInfo::
Evan Cheng4350eb82009-02-06 17:17:30 +0000[diff] [blame]4116isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4117 // FIXME: Return false for x87 stack register classes for now. We can't
Evan Cheng23066282008-10-27 07:14:50 +0000[diff] [blame]4118 // allow any loads of these registers before FpGet_ST0_80.
Evan Cheng4350eb82009-02-06 17:17:30 +0000[diff] [blame]4119 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4120 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
Evan Cheng23066282008-10-27 07:14:50 +0000[diff] [blame]4121}
4122
Dan Gohman57c3dac2008-09-30 00:58:23 +0000[diff] [blame]4123/// getGlobalBaseReg - Return a virtual register initialized with the
4124/// the global base register value. Output instructions required to
4125/// initialize the register in the function entry block, if necessary.
Dan Gohman8b746962008-09-23 18:22:58 +0000[diff] [blame]4126///
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4127/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4128///
Dan Gohman57c3dac2008-09-30 00:58:23 +0000[diff] [blame]4129unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4130 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4131 "X86-64 PIC uses RIP relative addressing");
4132
4133 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4134 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4135 if (GlobalBaseReg != 0)
4136 return GlobalBaseReg;
4137
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4138 // Create the register. The code to initialize it is inserted
4139 // later, by the CGBR pass (below).
Dan Gohman8b746962008-09-23 18:22:58 +0000[diff] [blame]4140 MachineRegisterInfo &RegInfo = MF->getRegInfo();
Jakob Stoklund Olesen53df9252012-05-20 18:43:00 +0000[diff] [blame]4141 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Dan Gohman57c3dac2008-09-30 00:58:23 +0000[diff] [blame]4142 X86FI->setGlobalBaseReg(GlobalBaseReg);
4143 return GlobalBaseReg;
Dan Gohman8b746962008-09-23 18:22:58 +0000[diff] [blame]4144}
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4145
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4146// These are the replaceable SSE instructions. Some of these have Int variants
4147// that we don't include here. We don't want to replace instructions selected
4148// by intrinsics.
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4149static const uint16_t ReplaceableInstrs[][3] = {
Bruno Cardoso Lopes4d043622010-08-12 02:08:52 +0000[diff] [blame]4150 //PackedSingle PackedDouble PackedInt
Jakob Stoklund Olesen357be7f2010-03-30 22:46:53 +0000[diff] [blame]4151 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4152 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4153 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4154 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4155 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4156 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4157 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4158 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4159 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4160 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4161 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4162 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4163 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4164 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]4165 // AVX 128-bit support
4166 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4167 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4168 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4169 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4170 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4171 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4172 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4173 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4174 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4175 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4176 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4177 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]4178 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4179 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]4180 // AVX 256-bit support
4181 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4182 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4183 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4184 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4185 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
Craig Topper4c077a12011-11-15 05:55:35 +0000[diff] [blame]4186 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4187};
4188
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4189static const uint16_t ReplaceableInstrsAVX2[][3] = {
Craig Topper4c077a12011-11-15 05:55:35 +0000[diff] [blame]4190 //PackedSingle PackedDouble PackedInt
Craig Topperb80ada92011-11-09 09:37:21 +0000[diff] [blame]4191 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4192 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4193 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4194 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4195 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4196 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4197 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
Craig Topperfe2a6c52011-11-29 05:37:58 +0000[diff] [blame]4198 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4199 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4200 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4201 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4202 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4203 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4204 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4205};
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4206
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4207// FIXME: Some shuffle and unpack instructions have equivalents in different
4208// domains, but they require a bit more work than just switching opcodes.
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4209
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4210static const uint16_t *lookup(unsigned opcode, unsigned domain) {
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4211 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4212 if (ReplaceableInstrs[i][domain-1] == opcode)
4213 return ReplaceableInstrs[i];
Craig Topper44ec9fd2011-11-15 06:39:01 +0000[diff] [blame]4214 return 0;
4215}
4216
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4217static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
Craig Topper44ec9fd2011-11-15 06:39:01 +0000[diff] [blame]4218 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4219 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4220 return ReplaceableInstrsAVX2[i];
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4221 return 0;
4222}
4223
4224std::pair<uint16_t, uint16_t>
Jakob Stoklund Olesen98e933f2011-09-27 22:57:18 +0000[diff] [blame]4225X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4226 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
Craig Topper4c077a12011-11-15 05:55:35 +0000[diff] [blame]4227 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
Craig Topper44ec9fd2011-11-15 06:39:01 +0000[diff] [blame]4228 uint16_t validDomains = 0;
4229 if (domain && lookup(MI->getOpcode(), domain))
4230 validDomains = 0xe;
4231 else if (domain && lookupAVX2(MI->getOpcode(), domain))
4232 validDomains = hasAVX2 ? 0xe : 0x6;
4233 return std::make_pair(domain, validDomains);
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4234}
4235
Jakob Stoklund Olesen98e933f2011-09-27 22:57:18 +0000[diff] [blame]4236void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4237 assert(Domain>0 && Domain<4 && "Invalid execution domain");
4238 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4239 assert(dom && "Not an SSE instruction");
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4240 const uint16_t *table = lookup(MI->getOpcode(), dom);
Jakob Stoklund Olesen7f5e43f2011-11-23 04:03:08 +0000[diff] [blame]4241 if (!table) { // try the other table
4242 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
4243 "256-bit vector operations only available in AVX2");
Craig Topper44ec9fd2011-11-15 06:39:01 +0000[diff] [blame]4244 table = lookupAVX2(MI->getOpcode(), dom);
Jakob Stoklund Olesen7f5e43f2011-11-23 04:03:08 +0000[diff] [blame]4245 }
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4246 assert(table && "Cannot change domain");
4247 MI->setDesc(get(table[Domain-1]));
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4248}
Chris Lattneree9eb412010-04-26 23:37:21 +0000[diff] [blame]4249
4250/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
4251void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
4252 NopInst.setOpcode(X86::NOOP);
4253}
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4254
Andrew Tricke0ef5092011-03-05 08:00:22 +0000[diff] [blame]4255bool X86InstrInfo::isHighLatencyDef(int opc) const {
4256 switch (opc) {
Evan Cheng23128422010-10-19 18:58:51 +0000[diff] [blame]4257 default: return false;
4258 case X86::DIVSDrm:
4259 case X86::DIVSDrm_Int:
4260 case X86::DIVSDrr:
4261 case X86::DIVSDrr_Int:
4262 case X86::DIVSSrm:
4263 case X86::DIVSSrm_Int:
4264 case X86::DIVSSrr:
4265 case X86::DIVSSrr_Int:
4266 case X86::SQRTPDm:
4267 case X86::SQRTPDm_Int:
4268 case X86::SQRTPDr:
4269 case X86::SQRTPDr_Int:
4270 case X86::SQRTPSm:
4271 case X86::SQRTPSm_Int:
4272 case X86::SQRTPSr:
4273 case X86::SQRTPSr_Int:
4274 case X86::SQRTSDm:
4275 case X86::SQRTSDm_Int:
4276 case X86::SQRTSDr:
4277 case X86::SQRTSDr_Int:
4278 case X86::SQRTSSm:
4279 case X86::SQRTSSm_Int:
4280 case X86::SQRTSSr:
4281 case X86::SQRTSSr_Int:
Bruno Cardoso Lopesb4e905d2011-09-15 22:15:52 +0000[diff] [blame]4282 // AVX instructions with high latency
4283 case X86::VDIVSDrm:
4284 case X86::VDIVSDrm_Int:
4285 case X86::VDIVSDrr:
4286 case X86::VDIVSDrr_Int:
4287 case X86::VDIVSSrm:
4288 case X86::VDIVSSrm_Int:
4289 case X86::VDIVSSrr:
4290 case X86::VDIVSSrr_Int:
4291 case X86::VSQRTPDm:
4292 case X86::VSQRTPDm_Int:
4293 case X86::VSQRTPDr:
4294 case X86::VSQRTPDr_Int:
4295 case X86::VSQRTPSm:
4296 case X86::VSQRTPSm_Int:
4297 case X86::VSQRTPSr:
4298 case X86::VSQRTPSr_Int:
4299 case X86::VSQRTSDm:
4300 case X86::VSQRTSDm_Int:
4301 case X86::VSQRTSDr:
4302 case X86::VSQRTSSm:
4303 case X86::VSQRTSSm_Int:
4304 case X86::VSQRTSSr:
Evan Cheng23128422010-10-19 18:58:51 +0000[diff] [blame]4305 return true;
4306 }
4307}
4308
Andrew Tricke0ef5092011-03-05 08:00:22 +0000[diff] [blame]4309bool X86InstrInfo::
4310hasHighOperandLatency(const InstrItineraryData *ItinData,
4311 const MachineRegisterInfo *MRI,
4312 const MachineInstr *DefMI, unsigned DefIdx,
4313 const MachineInstr *UseMI, unsigned UseIdx) const {
4314 return isHighLatencyDef(DefMI->getOpcode());
4315}
4316
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4317namespace {
4318 /// CGBR - Create Global Base Reg pass. This initializes the PIC
4319 /// global base register for x86-32.
4320 struct CGBR : public MachineFunctionPass {
4321 static char ID;
Owen Anderson90c579d2010-08-06 18:33:48 +0000[diff] [blame]4322 CGBR() : MachineFunctionPass(ID) {}
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4323
4324 virtual bool runOnMachineFunction(MachineFunction &MF) {
4325 const X86TargetMachine *TM =
4326 static_cast<const X86TargetMachine *>(&MF.getTarget());
4327
4328 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
4329 "X86-64 PIC uses RIP relative addressing");
4330
4331 // Only emit a global base reg in PIC mode.
4332 if (TM->getRelocationModel() != Reloc::PIC_)
4333 return false;
4334
Dan Gohmand8c0a512010-09-17 20:24:24 +0000[diff] [blame]4335 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
4336 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4337
4338 // If we didn't need a GlobalBaseReg, don't insert code.
4339 if (GlobalBaseReg == 0)
4340 return false;
4341
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4342 // Insert the set of GlobalBaseReg into the first MBB of the function
4343 MachineBasicBlock &FirstMBB = MF.front();
4344 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
4345 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
4346 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4347 const X86InstrInfo *TII = TM->getInstrInfo();
4348
4349 unsigned PC;
4350 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]4351 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4352 else
Dan Gohmand8c0a512010-09-17 20:24:24 +0000[diff] [blame]4353 PC = GlobalBaseReg;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]4354
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4355 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4356 // only used in JIT code emission as displacement to pc.
4357 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]4358
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4359 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4360 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4361 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4362 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4363 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4364 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4365 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4366 }
4367
4368 return true;
4369 }
4370
4371 virtual const char *getPassName() const {
4372 return "X86 PIC Global Base Reg Initialization";
4373 }
4374
4375 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4376 AU.setPreservesCFG();
4377 MachineFunctionPass::getAnalysisUsage(AU);
4378 }
4379 };
4380}
4381
4382char CGBR::ID = 0;
4383FunctionPass*
4384llvm::createGlobalBaseRegPass() { return new CGBR(); }
Hans Wennborgf0234fc2012-06-01 16:27:21 +0000[diff] [blame]4385
4386namespace {
4387 struct LDTLSCleanup : public MachineFunctionPass {
4388 static char ID;
4389 LDTLSCleanup() : MachineFunctionPass(ID) {}
4390
4391 virtual bool runOnMachineFunction(MachineFunction &MF) {
4392 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4393 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4394 // No point folding accesses if there isn't at least two.
4395 return false;
4396 }
4397
4398 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4399 return VisitNode(DT->getRootNode(), 0);
4400 }
4401
4402 // Visit the dominator subtree rooted at Node in pre-order.
4403 // If TLSBaseAddrReg is non-null, then use that to replace any
4404 // TLS_base_addr instructions. Otherwise, create the register
4405 // when the first such instruction is seen, and then use it
4406 // as we encounter more instructions.
4407 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4408 MachineBasicBlock *BB = Node->getBlock();
4409 bool Changed = false;
4410
4411 // Traverse the current block.
4412 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4413 ++I) {
4414 switch (I->getOpcode()) {
4415 case X86::TLS_base_addr32:
4416 case X86::TLS_base_addr64:
4417 if (TLSBaseAddrReg)
4418 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4419 else
4420 I = SetRegister(I, &TLSBaseAddrReg);
4421 Changed = true;
4422 break;
4423 default:
4424 break;
4425 }
4426 }
4427
4428 // Visit the children of this block in the dominator tree.
4429 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4430 I != E; ++I) {
4431 Changed |= VisitNode(*I, TLSBaseAddrReg);
4432 }
4433
4434 return Changed;
4435 }
4436
4437 // Replace the TLS_base_addr instruction I with a copy from
4438 // TLSBaseAddrReg, returning the new instruction.
4439 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4440 unsigned TLSBaseAddrReg) {
4441 MachineFunction *MF = I->getParent()->getParent();
4442 const X86TargetMachine *TM =
4443 static_cast<const X86TargetMachine *>(&MF->getTarget());
4444 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4445 const X86InstrInfo *TII = TM->getInstrInfo();
4446
4447 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4448 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4449 TII->get(TargetOpcode::COPY),
4450 is64Bit ? X86::RAX : X86::EAX)
4451 .addReg(TLSBaseAddrReg);
4452
4453 // Erase the TLS_base_addr instruction.
4454 I->eraseFromParent();
4455
4456 return Copy;
4457 }
4458
4459 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4460 // inserting a copy instruction after I. Returns the new instruction.
4461 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4462 MachineFunction *MF = I->getParent()->getParent();
4463 const X86TargetMachine *TM =
4464 static_cast<const X86TargetMachine *>(&MF->getTarget());
4465 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4466 const X86InstrInfo *TII = TM->getInstrInfo();
4467
4468 // Create a virtual register for the TLS base address.
4469 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4470 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4471 ? &X86::GR64RegClass
4472 : &X86::GR32RegClass);
4473
4474 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4475 MachineInstr *Next = I->getNextNode();
4476 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4477 TII->get(TargetOpcode::COPY),
4478 *TLSBaseAddrReg)
4479 .addReg(is64Bit ? X86::RAX : X86::EAX);
4480
4481 return Copy;
4482 }
4483
4484 virtual const char *getPassName() const {
4485 return "Local Dynamic TLS Access Clean-up";
4486 }
4487
4488 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4489 AU.setPreservesCFG();
4490 AU.addRequired<MachineDominatorTree>();
4491 MachineFunctionPass::getAnalysisUsage(AU);
4492 }
4493 };
4494}
4495
4496char LDTLSCleanup::ID = 0;
4497FunctionPass*
4498llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }