Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / 375d6c1ee033e8ce69fec9a9069cd80aa49a8f3a / . / llvm / lib / Target / X86 / X86InstrInfo.cpp
blob: 80d681a13c64b55f4dc004e74846df9a8b250ad7 [file] [log] [blame]
Jia Liub22310f2012-02-18 12:03:15 +0000[diff] [blame]1//===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
Misha Brukmanc88330a2005-04-21 23:38:14 +0000[diff] [blame]2//
John Criswell482202a2003-10-20 19:43:21 +0000[diff] [blame]3// The LLVM Compiler Infrastructure
4//
Chris Lattnerf3ebc3f2007-12-29 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 Brukmanc88330a2005-04-21 23:38:14 +0000[diff] [blame]7//
John Criswell482202a2003-10-20 19:43:21 +0000[diff] [blame]8//===----------------------------------------------------------------------===//
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]9//
Chris Lattnerb4d58d72003-01-14 22:00:31 +0000[diff] [blame]10// This file contains the X86 implementation of the TargetInstrInfo class.
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]11//
12//===----------------------------------------------------------------------===//
13
Chris Lattner27d24792002-10-29 21:05:24 +0000[diff] [blame]14#include "X86InstrInfo.h"
Chris Lattner0d808742002-12-03 05:42:53 +0000[diff] [blame]15#include "X86.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]16#include "X86InstrBuilder.h"
Owen Anderson6bb0c522008-01-04 23:57:37 +0000[diff] [blame]17#include "X86MachineFunctionInfo.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]18#include "X86Subtarget.h"
19#include "X86TargetMachine.h"
Owen Andersone2f23a32007-09-07 04:06:50 +0000[diff] [blame]20#include "llvm/ADT/STLExtras.h"
Chandler Carruthed0881b2012-12-03 16:50:05 +0000[diff] [blame]21#include "llvm/CodeGen/LiveVariables.h"
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]22#include "llvm/CodeGen/MachineConstantPool.h"
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]23#include "llvm/CodeGen/MachineDominators.h"
Owen Anderson6bb0c522008-01-04 23:57:37 +0000[diff] [blame]24#include "llvm/CodeGen/MachineFrameInfo.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]25#include "llvm/CodeGen/MachineInstrBuilder.h"
Chris Lattnera10fff52007-12-31 04:13:23 +0000[diff] [blame]26#include "llvm/CodeGen/MachineRegisterInfo.h"
Chandler Carruth9fb823b2013-01-02 11:36:10 +0000[diff] [blame]27#include "llvm/IR/DerivedTypes.h"
28#include "llvm/IR/LLVMContext.h"
Craig Topperb25fda92012-03-17 18:46:09 +0000[diff] [blame]29#include "llvm/MC/MCAsmInfo.h"
Chris Lattner6a5e7062010-04-26 23:37:21 +0000[diff] [blame]30#include "llvm/MC/MCInst.h"
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]31#include "llvm/Support/CommandLine.h"
David Greened589daf2010-01-05 01:29:29 +0000[diff] [blame]32#include "llvm/Support/Debug.h"
Torok Edwin6dd27302009-07-08 18:01:40 +0000[diff] [blame]33#include "llvm/Support/ErrorHandling.h"
34#include "llvm/Support/raw_ostream.h"
Evan Chenge95f3912007-09-25 01:57:46 +0000[diff] [blame]35#include "llvm/Target/TargetOptions.h"
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]36#include <limits>
37
Evan Cheng703a0fb2011-07-01 17:57:27 +0000[diff] [blame]38#define GET_INSTRINFO_CTOR
Evan Cheng1e210d02011-06-28 20:07:07 +0000[diff] [blame]39#include "X86GenInstrInfo.inc"
40
Brian Gaeke960707c2003-11-11 22:41:34 +0000[diff] [blame]41using namespace llvm;
42
Chris Lattnera6f074f2009-08-23 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 Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]55
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]56enum {
57 // Select which memory operand is being unfolded.
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]58 // (stored in bits 0 - 3)
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]59 TB_INDEX_0 = 0,
60 TB_INDEX_1 = 1,
61 TB_INDEX_2 = 2,
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]62 TB_INDEX_3 = 3,
Craig Topper1cac50b2012-06-23 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 Lopes23eb5262011-09-08 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,
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]84 TB_ALIGN_64 = 64 << TB_ALIGN_SHIFT,
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]85 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]86};
87
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]88struct X86OpTblEntry {
89 uint16_t RegOp;
90 uint16_t MemOp;
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]91 uint16_t Flags;
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]92};
93
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]94X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
Evan Cheng703a0fb2011-07-01 17:57:27 +0000[diff] [blame]95 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
96 ? X86::ADJCALLSTACKDOWN64
97 : X86::ADJCALLSTACKDOWN32),
98 (tm.getSubtarget<X86Subtarget>().is64Bit()
99 ? X86::ADJCALLSTACKUP64
100 : X86::ADJCALLSTACKUP32)),
Bill Wendling8f268402013-06-07 21:00:34 +0000[diff] [blame]101 TM(tm), RI(tm) {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]102
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]103 static const X86OpTblEntry OpTbl2Addr[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]104 { X86::ADC32ri, X86::ADC32mi, 0 },
105 { X86::ADC32ri8, X86::ADC32mi8, 0 },
106 { X86::ADC32rr, X86::ADC32mr, 0 },
107 { X86::ADC64ri32, X86::ADC64mi32, 0 },
108 { X86::ADC64ri8, X86::ADC64mi8, 0 },
109 { X86::ADC64rr, X86::ADC64mr, 0 },
110 { X86::ADD16ri, X86::ADD16mi, 0 },
111 { X86::ADD16ri8, X86::ADD16mi8, 0 },
112 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
113 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
114 { X86::ADD16rr, X86::ADD16mr, 0 },
115 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
116 { X86::ADD32ri, X86::ADD32mi, 0 },
117 { X86::ADD32ri8, X86::ADD32mi8, 0 },
118 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
119 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
120 { X86::ADD32rr, X86::ADD32mr, 0 },
121 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
122 { X86::ADD64ri32, X86::ADD64mi32, 0 },
123 { X86::ADD64ri8, X86::ADD64mi8, 0 },
124 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
125 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
126 { X86::ADD64rr, X86::ADD64mr, 0 },
127 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
128 { X86::ADD8ri, X86::ADD8mi, 0 },
129 { X86::ADD8rr, X86::ADD8mr, 0 },
130 { X86::AND16ri, X86::AND16mi, 0 },
131 { X86::AND16ri8, X86::AND16mi8, 0 },
132 { X86::AND16rr, X86::AND16mr, 0 },
133 { X86::AND32ri, X86::AND32mi, 0 },
134 { X86::AND32ri8, X86::AND32mi8, 0 },
135 { X86::AND32rr, X86::AND32mr, 0 },
136 { X86::AND64ri32, X86::AND64mi32, 0 },
137 { X86::AND64ri8, X86::AND64mi8, 0 },
138 { X86::AND64rr, X86::AND64mr, 0 },
139 { X86::AND8ri, X86::AND8mi, 0 },
140 { X86::AND8rr, X86::AND8mr, 0 },
141 { X86::DEC16r, X86::DEC16m, 0 },
142 { X86::DEC32r, X86::DEC32m, 0 },
143 { X86::DEC64_16r, X86::DEC64_16m, 0 },
144 { X86::DEC64_32r, X86::DEC64_32m, 0 },
145 { X86::DEC64r, X86::DEC64m, 0 },
146 { X86::DEC8r, X86::DEC8m, 0 },
147 { X86::INC16r, X86::INC16m, 0 },
148 { X86::INC32r, X86::INC32m, 0 },
149 { X86::INC64_16r, X86::INC64_16m, 0 },
150 { X86::INC64_32r, X86::INC64_32m, 0 },
151 { X86::INC64r, X86::INC64m, 0 },
152 { X86::INC8r, X86::INC8m, 0 },
153 { X86::NEG16r, X86::NEG16m, 0 },
154 { X86::NEG32r, X86::NEG32m, 0 },
155 { X86::NEG64r, X86::NEG64m, 0 },
156 { X86::NEG8r, X86::NEG8m, 0 },
157 { X86::NOT16r, X86::NOT16m, 0 },
158 { X86::NOT32r, X86::NOT32m, 0 },
159 { X86::NOT64r, X86::NOT64m, 0 },
160 { X86::NOT8r, X86::NOT8m, 0 },
161 { X86::OR16ri, X86::OR16mi, 0 },
162 { X86::OR16ri8, X86::OR16mi8, 0 },
163 { X86::OR16rr, X86::OR16mr, 0 },
164 { X86::OR32ri, X86::OR32mi, 0 },
165 { X86::OR32ri8, X86::OR32mi8, 0 },
166 { X86::OR32rr, X86::OR32mr, 0 },
167 { X86::OR64ri32, X86::OR64mi32, 0 },
168 { X86::OR64ri8, X86::OR64mi8, 0 },
169 { X86::OR64rr, X86::OR64mr, 0 },
170 { X86::OR8ri, X86::OR8mi, 0 },
171 { X86::OR8rr, X86::OR8mr, 0 },
172 { X86::ROL16r1, X86::ROL16m1, 0 },
173 { X86::ROL16rCL, X86::ROL16mCL, 0 },
174 { X86::ROL16ri, X86::ROL16mi, 0 },
175 { X86::ROL32r1, X86::ROL32m1, 0 },
176 { X86::ROL32rCL, X86::ROL32mCL, 0 },
177 { X86::ROL32ri, X86::ROL32mi, 0 },
178 { X86::ROL64r1, X86::ROL64m1, 0 },
179 { X86::ROL64rCL, X86::ROL64mCL, 0 },
180 { X86::ROL64ri, X86::ROL64mi, 0 },
181 { X86::ROL8r1, X86::ROL8m1, 0 },
182 { X86::ROL8rCL, X86::ROL8mCL, 0 },
183 { X86::ROL8ri, X86::ROL8mi, 0 },
184 { X86::ROR16r1, X86::ROR16m1, 0 },
185 { X86::ROR16rCL, X86::ROR16mCL, 0 },
186 { X86::ROR16ri, X86::ROR16mi, 0 },
187 { X86::ROR32r1, X86::ROR32m1, 0 },
188 { X86::ROR32rCL, X86::ROR32mCL, 0 },
189 { X86::ROR32ri, X86::ROR32mi, 0 },
190 { X86::ROR64r1, X86::ROR64m1, 0 },
191 { X86::ROR64rCL, X86::ROR64mCL, 0 },
192 { X86::ROR64ri, X86::ROR64mi, 0 },
193 { X86::ROR8r1, X86::ROR8m1, 0 },
194 { X86::ROR8rCL, X86::ROR8mCL, 0 },
195 { X86::ROR8ri, X86::ROR8mi, 0 },
196 { X86::SAR16r1, X86::SAR16m1, 0 },
197 { X86::SAR16rCL, X86::SAR16mCL, 0 },
198 { X86::SAR16ri, X86::SAR16mi, 0 },
199 { X86::SAR32r1, X86::SAR32m1, 0 },
200 { X86::SAR32rCL, X86::SAR32mCL, 0 },
201 { X86::SAR32ri, X86::SAR32mi, 0 },
202 { X86::SAR64r1, X86::SAR64m1, 0 },
203 { X86::SAR64rCL, X86::SAR64mCL, 0 },
204 { X86::SAR64ri, X86::SAR64mi, 0 },
205 { X86::SAR8r1, X86::SAR8m1, 0 },
206 { X86::SAR8rCL, X86::SAR8mCL, 0 },
207 { X86::SAR8ri, X86::SAR8mi, 0 },
208 { X86::SBB32ri, X86::SBB32mi, 0 },
209 { X86::SBB32ri8, X86::SBB32mi8, 0 },
210 { X86::SBB32rr, X86::SBB32mr, 0 },
211 { X86::SBB64ri32, X86::SBB64mi32, 0 },
212 { X86::SBB64ri8, X86::SBB64mi8, 0 },
213 { X86::SBB64rr, X86::SBB64mr, 0 },
214 { X86::SHL16rCL, X86::SHL16mCL, 0 },
215 { X86::SHL16ri, X86::SHL16mi, 0 },
216 { X86::SHL32rCL, X86::SHL32mCL, 0 },
217 { X86::SHL32ri, X86::SHL32mi, 0 },
218 { X86::SHL64rCL, X86::SHL64mCL, 0 },
219 { X86::SHL64ri, X86::SHL64mi, 0 },
220 { X86::SHL8rCL, X86::SHL8mCL, 0 },
221 { X86::SHL8ri, X86::SHL8mi, 0 },
222 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
223 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
224 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
225 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
226 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
227 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
228 { X86::SHR16r1, X86::SHR16m1, 0 },
229 { X86::SHR16rCL, X86::SHR16mCL, 0 },
230 { X86::SHR16ri, X86::SHR16mi, 0 },
231 { X86::SHR32r1, X86::SHR32m1, 0 },
232 { X86::SHR32rCL, X86::SHR32mCL, 0 },
233 { X86::SHR32ri, X86::SHR32mi, 0 },
234 { X86::SHR64r1, X86::SHR64m1, 0 },
235 { X86::SHR64rCL, X86::SHR64mCL, 0 },
236 { X86::SHR64ri, X86::SHR64mi, 0 },
237 { X86::SHR8r1, X86::SHR8m1, 0 },
238 { X86::SHR8rCL, X86::SHR8mCL, 0 },
239 { X86::SHR8ri, X86::SHR8mi, 0 },
240 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
241 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
242 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
243 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
244 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
245 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
246 { X86::SUB16ri, X86::SUB16mi, 0 },
247 { X86::SUB16ri8, X86::SUB16mi8, 0 },
248 { X86::SUB16rr, X86::SUB16mr, 0 },
249 { X86::SUB32ri, X86::SUB32mi, 0 },
250 { X86::SUB32ri8, X86::SUB32mi8, 0 },
251 { X86::SUB32rr, X86::SUB32mr, 0 },
252 { X86::SUB64ri32, X86::SUB64mi32, 0 },
253 { X86::SUB64ri8, X86::SUB64mi8, 0 },
254 { X86::SUB64rr, X86::SUB64mr, 0 },
255 { X86::SUB8ri, X86::SUB8mi, 0 },
256 { X86::SUB8rr, X86::SUB8mr, 0 },
257 { X86::XOR16ri, X86::XOR16mi, 0 },
258 { X86::XOR16ri8, X86::XOR16mi8, 0 },
259 { X86::XOR16rr, X86::XOR16mr, 0 },
260 { X86::XOR32ri, X86::XOR32mi, 0 },
261 { X86::XOR32ri8, X86::XOR32mi8, 0 },
262 { X86::XOR32rr, X86::XOR32mr, 0 },
263 { X86::XOR64ri32, X86::XOR64mi32, 0 },
264 { X86::XOR64ri8, X86::XOR64mi8, 0 },
265 { X86::XOR64rr, X86::XOR64mr, 0 },
266 { X86::XOR8ri, X86::XOR8mi, 0 },
267 { X86::XOR8rr, X86::XOR8mr, 0 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]268 };
269
270 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]271 unsigned RegOp = OpTbl2Addr[i].RegOp;
272 unsigned MemOp = OpTbl2Addr[i].MemOp;
273 unsigned Flags = OpTbl2Addr[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]274 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
275 RegOp, MemOp,
276 // Index 0, folded load and store, no alignment requirement.
277 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]278 }
279
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]280 static const X86OpTblEntry OpTbl0[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]281 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
282 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
283 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
284 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
285 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]286 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
287 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
288 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
289 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
290 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
291 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
292 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
293 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
294 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
295 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
296 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
297 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
298 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
299 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
300 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
Craig Topperd09a9af2012-12-26 01:47:12 +0000[diff] [blame]301 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]302 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
303 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]304 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
305 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
306 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
307 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
308 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
309 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
310 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
311 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
312 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
313 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
314 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
315 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
316 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
317 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
318 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
319 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
320 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
321 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
322 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
323 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
324 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
325 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]326 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
327 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
328 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
329 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
330 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
331 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]332 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
333 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
334 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
335 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
336 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
337 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
338 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
339 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
340 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
341 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
342 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
343 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
344 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
345 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
346 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
347 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
348 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
349 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
350 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
351 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
352 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
353 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
354 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
355 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
356 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]357 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
358 // AVX 128-bit versions of foldable instructions
Craig Topperd09a9af2012-12-26 01:47:12 +0000[diff] [blame]359 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]360 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
361 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]362 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]363 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
366 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
367 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
368 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
369 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
370 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
371 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
372 // AVX 256-bit foldable instructions
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]373 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]374 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
377 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
Elena Demikhovsky534015e2013-09-02 07:12:29 +0000[diff] [blame]378 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE },
379 // AVX-512 foldable instructions
380 { X86::VMOVPDI2DIZrr,X86::VMOVPDI2DIZmr, TB_FOLDED_STORE }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]381 };
382
383 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]384 unsigned RegOp = OpTbl0[i].RegOp;
385 unsigned MemOp = OpTbl0[i].MemOp;
386 unsigned Flags = OpTbl0[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]387 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
388 RegOp, MemOp, TB_INDEX_0 | Flags);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]389 }
390
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]391 static const X86OpTblEntry OpTbl1[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]392 { X86::CMP16rr, X86::CMP16rm, 0 },
393 { X86::CMP32rr, X86::CMP32rm, 0 },
394 { X86::CMP64rr, X86::CMP64rm, 0 },
395 { X86::CMP8rr, X86::CMP8rm, 0 },
396 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
397 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
398 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
399 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
400 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
401 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
402 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
403 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
404 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
405 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
406 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
407 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]408 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
409 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
410 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
411 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
412 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
413 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
414 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
415 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]416 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
417 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]418 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
419 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]420 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
421 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
422 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
423 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
424 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
425 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
426 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
427 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]428 { X86::MOV16rr, X86::MOV16rm, 0 },
429 { X86::MOV32rr, X86::MOV32rm, 0 },
430 { X86::MOV64rr, X86::MOV64rm, 0 },
431 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
432 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
433 { X86::MOV8rr, X86::MOV8rm, 0 },
434 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
435 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]436 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
437 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
438 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
439 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]440 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
441 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
442 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
443 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
444 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
445 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
446 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
447 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
448 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
449 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]450 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
451 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
452 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
453 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
454 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
455 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
456 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]457 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
458 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
459 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]460 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
461 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
462 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
463 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
464 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
465 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
466 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
467 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
468 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
469 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]470 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]471 { X86::SQRTSDr, X86::SQRTSDm, 0 },
472 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
473 { X86::SQRTSSr, X86::SQRTSSm, 0 },
474 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
475 { X86::TEST16rr, X86::TEST16rm, 0 },
476 { X86::TEST32rr, X86::TEST32rm, 0 },
477 { X86::TEST64rr, X86::TEST64rm, 0 },
478 { X86::TEST8rr, X86::TEST8rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]479 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]480 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
481 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]482 // AVX 128-bit versions of foldable instructions
483 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
484 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]485 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
486 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]487 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
488 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
Pete Cooper8bbce762012-06-14 22:12:58 +0000[diff] [blame]489 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]490 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
491 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
492 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
493 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
494 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
495 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
496 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
497 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
498 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]499 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
500 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
501 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
502 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
503 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
504 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
505 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
506 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
507 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
508 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
509 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
510 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
Craig Topperb2922162012-12-26 02:14:19 +0000[diff] [blame]511 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]512 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
513 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
514 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
515 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]516 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
517 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
518 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
519 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
520 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
521 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
522 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
523 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
524 { X86::VRCPPSr, X86::VRCPPSm, 0 },
525 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, 0 },
526 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
527 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, 0 },
528 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]529 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]530 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]531 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]532 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
533
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]534 // AVX 256-bit foldable instructions
535 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
536 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]537 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]538 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]539 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]540 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
541 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]542
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]543 // AVX2 foldable instructions
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]544 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
545 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
546 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
547 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
548 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
549 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
550 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
551 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, 0 },
552 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]553 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]554 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]555 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
556 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
Michael Liao2de86af2012-09-26 08:24:51 +0000[diff] [blame]557
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]558 // BMI/BMI2/LZCNT/POPCNT foldable instructions
559 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
560 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
561 { X86::BLSI32rr, X86::BLSI32rm, 0 },
562 { X86::BLSI64rr, X86::BLSI64rm, 0 },
563 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
564 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
565 { X86::BLSR32rr, X86::BLSR32rm, 0 },
566 { X86::BLSR64rr, X86::BLSR64rm, 0 },
567 { X86::BZHI32rr, X86::BZHI32rm, 0 },
568 { X86::BZHI64rr, X86::BZHI64rm, 0 },
569 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
570 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
571 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
572 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
573 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
574 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
Michael Liao2de86af2012-09-26 08:24:51 +0000[diff] [blame]575 { X86::RORX32ri, X86::RORX32mi, 0 },
576 { X86::RORX64ri, X86::RORX64mi, 0 },
Michael Liao2b425e12012-09-26 08:26:25 +0000[diff] [blame]577 { X86::SARX32rr, X86::SARX32rm, 0 },
578 { X86::SARX64rr, X86::SARX64rm, 0 },
579 { X86::SHRX32rr, X86::SHRX32rm, 0 },
580 { X86::SHRX64rr, X86::SHRX64rm, 0 },
581 { X86::SHLX32rr, X86::SHLX32rm, 0 },
582 { X86::SHLX64rr, X86::SHLX64rm, 0 },
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]583 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
584 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
585 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
Elena Demikhovsky534015e2013-09-02 07:12:29 +0000[diff] [blame]586
587 // AVX-512 foldable instructions
588 { X86::VMOV64toPQIZrr, X86::VMOVQI2PQIZrm, 0 },
589 { X86::VMOVDI2SSZrr, X86::VMOVDI2SSZrm, 0 },
Craig Topper684abc82013-09-17 06:05:17 +0000[diff] [blame]590 { X86::VMOVDQA32rr, X86::VMOVDQA32rm, TB_ALIGN_64 },
591 { X86::VMOVDQA64rr, X86::VMOVDQA64rm, TB_ALIGN_64 },
592 { X86::VMOVDQU32rr, X86::VMOVDQU32rm, 0 },
593 { X86::VMOVDQU64rr, X86::VMOVDQU64rm, 0 },
Craig Topper514f02c2013-09-17 06:50:11 +0000[diff] [blame]594
595 // AES foldable instructions
596 { X86::AESIMCrr, X86::AESIMCrm, TB_ALIGN_16 },
597 { X86::AESKEYGENASSIST128rr, X86::AESKEYGENASSIST128rm, TB_ALIGN_16 },
598 { X86::VAESIMCrr, X86::VAESIMCrm, TB_ALIGN_16 },
599 { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, TB_ALIGN_16 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]600 };
601
602 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]603 unsigned RegOp = OpTbl1[i].RegOp;
604 unsigned MemOp = OpTbl1[i].MemOp;
605 unsigned Flags = OpTbl1[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]606 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
607 RegOp, MemOp,
608 // Index 1, folded load
609 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]610 }
611
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]612 static const X86OpTblEntry OpTbl2[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]613 { X86::ADC32rr, X86::ADC32rm, 0 },
614 { X86::ADC64rr, X86::ADC64rm, 0 },
615 { X86::ADD16rr, X86::ADD16rm, 0 },
616 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
617 { X86::ADD32rr, X86::ADD32rm, 0 },
618 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
619 { X86::ADD64rr, X86::ADD64rm, 0 },
620 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
621 { X86::ADD8rr, X86::ADD8rm, 0 },
622 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
623 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
624 { X86::ADDSDrr, X86::ADDSDrm, 0 },
625 { X86::ADDSSrr, X86::ADDSSrm, 0 },
626 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
627 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
628 { X86::AND16rr, X86::AND16rm, 0 },
629 { X86::AND32rr, X86::AND32rm, 0 },
630 { X86::AND64rr, X86::AND64rm, 0 },
631 { X86::AND8rr, X86::AND8rm, 0 },
632 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
633 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
634 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
635 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]636 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
637 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
638 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
639 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]640 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
641 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
642 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
643 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
644 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
645 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
646 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
647 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
648 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
649 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
650 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
651 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
652 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
653 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
654 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
655 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
656 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
657 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
658 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
659 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
660 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
661 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
662 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
663 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
664 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
665 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
666 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
667 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
668 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
669 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
670 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
671 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
672 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
673 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
674 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
675 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
676 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
677 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
678 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
679 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
680 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
681 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
682 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
683 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
684 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
685 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
686 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
687 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
688 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
689 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
690 { X86::CMPSDrr, X86::CMPSDrm, 0 },
691 { X86::CMPSSrr, X86::CMPSSrm, 0 },
692 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
693 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
694 { X86::DIVSDrr, X86::DIVSDrm, 0 },
695 { X86::DIVSSrr, X86::DIVSSrm, 0 },
696 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
697 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
698 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
699 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
700 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
701 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
702 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
703 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
704 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
705 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
706 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
707 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
708 { X86::IMUL16rr, X86::IMUL16rm, 0 },
709 { X86::IMUL32rr, X86::IMUL32rm, 0 },
710 { X86::IMUL64rr, X86::IMUL64rm, 0 },
711 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
712 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
Manman Ren959acb12012-08-13 18:29:41 +0000[diff] [blame]713 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
714 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
715 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
716 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
717 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
718 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]719 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]720 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]721 { X86::MAXSDrr, X86::MAXSDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]722 { X86::MAXSSrr, X86::MAXSSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]723 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]724 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]725 { X86::MINSDrr, X86::MINSDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]726 { X86::MINSSrr, X86::MINSSrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]727 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]728 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
729 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
730 { X86::MULSDrr, X86::MULSDrm, 0 },
731 { X86::MULSSrr, X86::MULSSrm, 0 },
732 { X86::OR16rr, X86::OR16rm, 0 },
733 { X86::OR32rr, X86::OR32rm, 0 },
734 { X86::OR64rr, X86::OR64rm, 0 },
735 { X86::OR8rr, X86::OR8rm, 0 },
736 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
737 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
738 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
739 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]740 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]741 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
742 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
743 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
744 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
745 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
746 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]747 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
748 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]749 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]750 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]751 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
752 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
753 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
754 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]755 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]756 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
757 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]758 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]759 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
760 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
761 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]762 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]763 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]764 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
765 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]766 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]767 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]768 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]769 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]770 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]771 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]772 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
773 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
774 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
775 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
776 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
Benjamin Kramer4669d182012-12-21 14:04:55 +0000[diff] [blame]777 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
778 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
779 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
780 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
781 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
782 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
783 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
784 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]785 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]786 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]787 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
788 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
789 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
790 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
791 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
792 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
793 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]794 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
795 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
796 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
797 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]798 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
799 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
800 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
801 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
802 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
803 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
804 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
805 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
806 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
807 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
808 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
809 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
810 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
811 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
812 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
813 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
814 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
815 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
816 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
817 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
818 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
819 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
820 { X86::SBB32rr, X86::SBB32rm, 0 },
821 { X86::SBB64rr, X86::SBB64rm, 0 },
822 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
823 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
824 { X86::SUB16rr, X86::SUB16rm, 0 },
825 { X86::SUB32rr, X86::SUB32rm, 0 },
826 { X86::SUB64rr, X86::SUB64rm, 0 },
827 { X86::SUB8rr, X86::SUB8rm, 0 },
828 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
829 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
830 { X86::SUBSDrr, X86::SUBSDrm, 0 },
831 { X86::SUBSSrr, X86::SUBSSrm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]832 // FIXME: TEST*rr -> swapped operand of TEST*mr.
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]833 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
834 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
835 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
836 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
837 { X86::XOR16rr, X86::XOR16rm, 0 },
838 { X86::XOR32rr, X86::XOR32rm, 0 },
839 { X86::XOR64rr, X86::XOR64rm, 0 },
840 { X86::XOR8rr, X86::XOR8rm, 0 },
841 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]842 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
843 // AVX 128-bit versions of foldable instructions
844 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
845 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
846 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
847 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
848 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
849 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
850 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
851 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
852 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
853 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
Craig Toppercaef1c52012-12-26 00:35:47 +0000[diff] [blame]854 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
855 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]856 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
857 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]858 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
859 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
860 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]861 { X86::VADDPDrr, X86::VADDPDrm, 0 },
862 { X86::VADDPSrr, X86::VADDPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]863 { X86::VADDSDrr, X86::VADDSDrm, 0 },
864 { X86::VADDSSrr, X86::VADDSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]865 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
866 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
867 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
868 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
869 { X86::VANDPDrr, X86::VANDPDrm, 0 },
870 { X86::VANDPSrr, X86::VANDPSrm, 0 },
871 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
872 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
873 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
874 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
875 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
876 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]877 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
878 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]879 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
880 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]881 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
882 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
883 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
884 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
885 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
886 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
887 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
888 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
889 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
890 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]891 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
892 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
893 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
894 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]895 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
896 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]897 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]898 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]899 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]900 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]901 { X86::VMINPDrr, X86::VMINPDrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]902 { X86::VMINPSrr, X86::VMINPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]903 { X86::VMINSDrr, X86::VMINSDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]904 { X86::VMINSSrr, X86::VMINSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]905 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
906 { X86::VMULPDrr, X86::VMULPDrm, 0 },
907 { X86::VMULPSrr, X86::VMULPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]908 { X86::VMULSDrr, X86::VMULSDrm, 0 },
909 { X86::VMULSSrr, X86::VMULSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]910 { X86::VORPDrr, X86::VORPDrm, 0 },
911 { X86::VORPSrr, X86::VORPSrm, 0 },
912 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
913 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
914 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
915 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
916 { X86::VPADDBrr, X86::VPADDBrm, 0 },
917 { X86::VPADDDrr, X86::VPADDDrm, 0 },
918 { X86::VPADDQrr, X86::VPADDQrm, 0 },
919 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
920 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
921 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
922 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
923 { X86::VPADDWrr, X86::VPADDWrm, 0 },
924 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
925 { X86::VPANDNrr, X86::VPANDNrm, 0 },
926 { X86::VPANDrr, X86::VPANDrm, 0 },
927 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
928 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
929 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
930 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
931 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
932 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
933 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
934 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
935 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
936 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
937 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
938 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
939 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
940 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
941 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
942 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
943 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
944 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
945 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
946 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
947 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
948 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
949 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
950 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
951 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
952 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
953 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
954 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
955 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
956 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
957 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
958 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
959 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
960 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
961 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
962 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
963 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
964 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
965 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
966 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
967 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
968 { X86::VPORrr, X86::VPORrm, 0 },
969 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
970 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
971 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
972 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
973 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
974 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
975 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
976 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
977 { X86::VPSRADrr, X86::VPSRADrm, 0 },
978 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
979 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
980 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
981 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
982 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
983 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
984 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
985 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
986 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
987 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
988 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
989 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
990 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
991 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
992 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
993 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
994 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
995 { X86::VPXORrr, X86::VPXORrm, 0 },
996 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
997 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
998 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
999 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1000 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
1001 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1002 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
1003 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
1004 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
1005 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
1006 { X86::VXORPDrr, X86::VXORPDrm, 0 },
1007 { X86::VXORPSrr, X86::VXORPSrm, 0 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1008 // AVX 256-bit foldable instructions
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1009 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
1010 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
1011 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
1012 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
1013 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
1014 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
1015 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
1016 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1017 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1018 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1019 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1020 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1021 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1022 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1023 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1024 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1025 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1026 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1027 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1028 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1029 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1030 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1031 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1032 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1033 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1034 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1035 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1036 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1037 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1038 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1039 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1040 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1041 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1042 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1043 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1044 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1045 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1046 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1047 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1048 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1049 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1050 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1051 // AVX2 foldable instructions
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1052 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1053 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1054 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1055 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1056 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1057 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1058 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1059 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1060 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1061 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1062 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1063 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1064 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1065 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1066 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1067 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1068 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1069 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1070 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1071 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1072 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1073 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1074 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1075 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1076 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1077 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1078 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1079 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1080 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1081 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1082 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1083 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
1084 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1085 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
1086 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1087 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1088 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1089 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1090 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1091 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1092 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1093 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1094 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1095 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1096 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1097 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1098 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1099 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1100 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1101 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1102 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1103 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1104 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1105 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1106 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1107 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1108 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1109 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1110 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1111 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1112 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1113 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1114 { X86::VPORYrr, X86::VPORYrm, 0 },
1115 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1116 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1117 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1118 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1119 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1120 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1121 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1122 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1123 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1124 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1125 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1126 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1127 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1128 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1129 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1130 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1131 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1132 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1133 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1134 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1135 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1136 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1137 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1138 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1139 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1140 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1141 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1142 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1143 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1144 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1145 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1146 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1147 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1148 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1149 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1150 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1151 { X86::VPXORYrr, X86::VPXORYrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1152 // FIXME: add AVX 256-bit foldable instructions
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1153
1154 // FMA4 foldable patterns
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1155 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, 0 },
1156 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1157 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1158 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1159 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1160 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1161 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, 0 },
1162 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1163 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1164 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1165 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1166 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1167 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, 0 },
1168 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1169 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1170 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1171 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1172 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1173 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, 0 },
1174 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1175 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1176 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1177 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1178 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1179 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1180 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1181 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1182 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1183 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1184 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1185 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1186 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
Michael Liaof9f7b552012-09-26 08:22:37 +0000[diff] [blame]1187
1188 // BMI/BMI2 foldable instructions
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]1189 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1190 { X86::ANDN64rr, X86::ANDN64rm, 0 },
Michael Liaof9f7b552012-09-26 08:22:37 +0000[diff] [blame]1191 { X86::MULX32rr, X86::MULX32rm, 0 },
1192 { X86::MULX64rr, X86::MULX64rm, 0 },
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]1193 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1194 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1195 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1196 { X86::PEXT64rr, X86::PEXT64rm, 0 },
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]1197
1198 // AVX-512 foldable instructions
Elena Demikhovsky534015e2013-09-02 07:12:29 +0000[diff] [blame]1199 { X86::VPADDDZrr, X86::VPADDDZrm, 0 },
1200 { X86::VPADDQZrr, X86::VPADDQZrm, 0 },
1201 { X86::VADDPSZrr, X86::VADDPSZrm, 0 },
1202 { X86::VADDPDZrr, X86::VADDPDZrm, 0 },
1203 { X86::VSUBPSZrr, X86::VSUBPSZrm, 0 },
1204 { X86::VSUBPDZrr, X86::VSUBPDZrm, 0 },
1205 { X86::VMULPSZrr, X86::VMULPSZrm, 0 },
1206 { X86::VMULPDZrr, X86::VMULPDZrm, 0 },
1207 { X86::VDIVPSZrr, X86::VDIVPSZrm, 0 },
1208 { X86::VDIVPDZrr, X86::VDIVPDZrm, 0 },
1209 { X86::VMINPSZrr, X86::VMINPSZrm, 0 },
1210 { X86::VMINPDZrr, X86::VMINPDZrm, 0 },
1211 { X86::VMAXPSZrr, X86::VMAXPSZrm, 0 },
1212 { X86::VMAXPDZrr, X86::VMAXPDZrm, 0 },
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]1213 { X86::VPERMPDZri, X86::VPERMPDZmi, 0 },
1214 { X86::VPERMPSZrr, X86::VPERMPSZrm, 0 },
1215 { X86::VPERMI2Drr, X86::VPERMI2Drm, 0 },
1216 { X86::VPERMI2Qrr, X86::VPERMI2Qrm, 0 },
1217 { X86::VPERMI2PSrr, X86::VPERMI2PSrm, 0 },
1218 { X86::VPERMI2PDrr, X86::VPERMI2PDrm, 0 },
Elena Demikhovsky534015e2013-09-02 07:12:29 +0000[diff] [blame]1219 { X86::VPSLLVDZrr, X86::VPSLLVDZrm, 0 },
1220 { X86::VPSLLVQZrr, X86::VPSLLVQZrm, 0 },
1221 { X86::VPSRAVDZrr, X86::VPSRAVDZrm, 0 },
1222 { X86::VPSRLVDZrr, X86::VPSRLVDZrm, 0 },
1223 { X86::VPSRLVQZrr, X86::VPSRLVQZrm, 0 },
1224 { X86::VSHUFPDZrri, X86::VSHUFPDZrmi, 0 },
1225 { X86::VSHUFPSZrri, X86::VSHUFPSZrmi, 0 },
1226 { X86::VALIGNQrri, X86::VALIGNQrmi, 0 },
1227 { X86::VALIGNDrri, X86::VALIGNDrmi, 0 },
Craig Topper514f02c2013-09-17 06:50:11 +0000[diff] [blame]1228
1229 // AES foldable instructions
1230 { X86::AESDECLASTrr, X86::AESDECLASTrm, TB_ALIGN_16 },
1231 { X86::AESDECrr, X86::AESDECrm, TB_ALIGN_16 },
1232 { X86::AESENCLASTrr, X86::AESENCLASTrm, TB_ALIGN_16 },
1233 { X86::AESENCrr, X86::AESENCrm, TB_ALIGN_16 },
1234 { X86::VAESDECLASTrr, X86::VAESDECLASTrm, TB_ALIGN_16 },
1235 { X86::VAESDECrr, X86::VAESDECrm, TB_ALIGN_16 },
1236 { X86::VAESENCLASTrr, X86::VAESENCLASTrm, TB_ALIGN_16 },
1237 { X86::VAESENCrr, X86::VAESENCrm, TB_ALIGN_16 },
1238
1239 // SHA foldable instructions
1240 { X86::SHA1MSG1rr, X86::SHA1MSG1rm, TB_ALIGN_16 },
1241 { X86::SHA1MSG2rr, X86::SHA1MSG2rm, TB_ALIGN_16 },
1242 { X86::SHA1NEXTErr, X86::SHA1NEXTErm, TB_ALIGN_16 },
1243 { X86::SHA1RNDS4rri, X86::SHA1RNDS4rmi, TB_ALIGN_16 },
1244 { X86::SHA256MSG1rr, X86::SHA256MSG1rm, TB_ALIGN_16 },
1245 { X86::SHA256MSG2rr, X86::SHA256MSG2rm, TB_ALIGN_16 },
1246 { X86::SHA256RNDS2rr, X86::SHA256RNDS2rm, TB_ALIGN_16 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1247 };
1248
1249 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]1250 unsigned RegOp = OpTbl2[i].RegOp;
1251 unsigned MemOp = OpTbl2[i].MemOp;
1252 unsigned Flags = OpTbl2[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1253 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1254 RegOp, MemOp,
1255 // Index 2, folded load
1256 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1257 }
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1258
1259 static const X86OpTblEntry OpTbl3[] = {
1260 // FMA foldable instructions
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1261 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1262 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1263 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1264 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1265 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1266 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1267 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, 0 },
1268 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1269
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1270 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1271 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1272 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1273 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1274 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1275 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1276 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1277 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1278 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1279 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1280 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1281 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1282
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1283 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1284 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1285 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1286 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1287 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1288 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1289 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, 0 },
1290 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1291
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1292 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1293 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1294 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1295 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1296 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1297 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1298 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1299 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1300 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1301 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1302 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1303 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1304
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1305 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1306 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1307 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1308 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1309 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1310 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1311 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, 0 },
1312 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1313
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1314 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1315 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1316 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1317 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1318 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1319 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1320 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1321 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1322 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1323 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1324 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1325 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1326
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1327 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1328 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1329 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1330 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1331 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1332 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1333 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, 0 },
1334 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, 0 },
Craig Topper2e127b52012-06-01 05:48:39 +0000[diff] [blame]1335
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1336 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1337 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1338 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1339 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1340 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1341 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1342 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1343 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1344 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1345 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1346 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1347 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1348
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1349 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1350 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1351 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1352 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1353 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1354 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1355 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1356 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1357 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1358 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1359 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1360 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1361
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1362 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1363 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1364 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1365 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1366 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1367 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1368 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1369 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1370 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1371 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1372 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1373 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1374
1375 // FMA4 foldable patterns
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1376 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, 0 },
1377 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1378 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1379 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1380 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1381 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1382 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, 0 },
1383 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1384 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1385 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1386 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1387 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1388 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, 0 },
1389 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1390 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1391 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1392 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1393 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1394 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, 0 },
1395 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1396 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1397 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1398 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1399 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1400 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1401 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1402 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1403 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1404 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1405 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1406 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1407 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1408 };
1409
1410 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1411 unsigned RegOp = OpTbl3[i].RegOp;
1412 unsigned MemOp = OpTbl3[i].MemOp;
1413 unsigned Flags = OpTbl3[i].Flags;
1414 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1415 RegOp, MemOp,
1416 // Index 3, folded load
1417 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1418 }
1419
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]1420}
1421
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1422void
1423X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1424 MemOp2RegOpTableType &M2RTable,
1425 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1426 if ((Flags & TB_NO_FORWARD) == 0) {
1427 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1428 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1429 }
1430 if ((Flags & TB_NO_REVERSE) == 0) {
1431 assert(!M2RTable.count(MemOp) &&
1432 "Duplicated entries in unfolding maps?");
1433 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1434 }
1435}
1436
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1437bool
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1438X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1439 unsigned &SrcReg, unsigned &DstReg,
1440 unsigned &SubIdx) const {
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1441 switch (MI.getOpcode()) {
1442 default: break;
1443 case X86::MOVSX16rr8:
1444 case X86::MOVZX16rr8:
1445 case X86::MOVSX32rr8:
1446 case X86::MOVZX32rr8:
1447 case X86::MOVSX64rr8:
Evan Chengceb5a4e2010-01-13 08:01:32 +0000[diff] [blame]1448 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1449 // It's not always legal to reference the low 8-bit of the larger
1450 // register in 32-bit mode.
1451 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1452 case X86::MOVSX32rr16:
1453 case X86::MOVZX32rr16:
1454 case X86::MOVSX64rr16:
Tim Northover04eb4232013-05-30 10:43:18 +0000[diff] [blame]1455 case X86::MOVSX64rr32: {
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1456 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1457 // Be conservative.
1458 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1459 SrcReg = MI.getOperand(1).getReg();
1460 DstReg = MI.getOperand(0).getReg();
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1461 switch (MI.getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1462 default: llvm_unreachable("Unreachable!");
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1463 case X86::MOVSX16rr8:
1464 case X86::MOVZX16rr8:
1465 case X86::MOVSX32rr8:
1466 case X86::MOVZX32rr8:
1467 case X86::MOVSX64rr8:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1468 SubIdx = X86::sub_8bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1469 break;
1470 case X86::MOVSX32rr16:
1471 case X86::MOVZX32rr16:
1472 case X86::MOVSX64rr16:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1473 SubIdx = X86::sub_16bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1474 break;
1475 case X86::MOVSX64rr32:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1476 SubIdx = X86::sub_32bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1477 break;
1478 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1479 return true;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1480 }
1481 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1482 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1483}
1484
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1485/// isFrameOperand - Return true and the FrameIndex if the specified
1486/// operand and follow operands form a reference to the stack frame.
1487bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1488 int &FrameIndex) const {
1489 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1490 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1491 MI->getOperand(Op+1).getImm() == 1 &&
1492 MI->getOperand(Op+2).getReg() == 0 &&
1493 MI->getOperand(Op+3).getImm() == 0) {
1494 FrameIndex = MI->getOperand(Op).getIndex();
1495 return true;
1496 }
1497 return false;
1498}
1499
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1500static bool isFrameLoadOpcode(int Opcode) {
1501 switch (Opcode) {
David Blaikie46a9f012012-01-20 21:51:11 +0000[diff] [blame]1502 default:
1503 return false;
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1504 case X86::MOV8rm:
1505 case X86::MOV16rm:
1506 case X86::MOV32rm:
Evan Cheng11b0a5d2006-09-08 06:48:29 +0000[diff] [blame]1507 case X86::MOV64rm:
Dale Johannesen3d7008c2007-07-04 21:07:47 +0000[diff] [blame]1508 case X86::LD_Fp64m:
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1509 case X86::MOVSSrm:
1510 case X86::MOVSDrm:
Chris Lattnerbfc2c682006-04-18 16:44:51 +0000[diff] [blame]1511 case X86::MOVAPSrm:
1512 case X86::MOVAPDrm:
Dan Gohmanbdc0f8b2009-01-09 02:40:34 +0000[diff] [blame]1513 case X86::MOVDQArm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1514 case X86::VMOVSSrm:
1515 case X86::VMOVSDrm:
1516 case X86::VMOVAPSrm:
1517 case X86::VMOVAPDrm:
1518 case X86::VMOVDQArm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1519 case X86::VMOVAPSYrm:
1520 case X86::VMOVAPDYrm:
1521 case X86::VMOVDQAYrm:
Bill Wendlinge7b2a862007-04-03 06:00:37 +0000[diff] [blame]1522 case X86::MMX_MOVD64rm:
1523 case X86::MMX_MOVQ64rm:
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]1524 case X86::VMOVDQA32rm:
1525 case X86::VMOVDQA64rm:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1526 return true;
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1527 }
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1528}
1529
1530static bool isFrameStoreOpcode(int Opcode) {
1531 switch (Opcode) {
1532 default: break;
1533 case X86::MOV8mr:
1534 case X86::MOV16mr:
1535 case X86::MOV32mr:
1536 case X86::MOV64mr:
1537 case X86::ST_FpP64m:
1538 case X86::MOVSSmr:
1539 case X86::MOVSDmr:
1540 case X86::MOVAPSmr:
1541 case X86::MOVAPDmr:
1542 case X86::MOVDQAmr:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1543 case X86::VMOVSSmr:
1544 case X86::VMOVSDmr:
1545 case X86::VMOVAPSmr:
1546 case X86::VMOVAPDmr:
1547 case X86::VMOVDQAmr:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1548 case X86::VMOVAPSYmr:
1549 case X86::VMOVAPDYmr:
1550 case X86::VMOVDQAYmr:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1551 case X86::MMX_MOVD64mr:
1552 case X86::MMX_MOVQ64mr:
1553 case X86::MMX_MOVNTQmr:
1554 return true;
1555 }
1556 return false;
1557}
1558
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1559unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1560 int &FrameIndex) const {
1561 if (isFrameLoadOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1562 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1563 return MI->getOperand(0).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1564 return 0;
1565}
1566
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1567unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1568 int &FrameIndex) const {
1569 if (isFrameLoadOpcode(MI->getOpcode())) {
1570 unsigned Reg;
1571 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1572 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1573 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1574 const MachineMemOperand *Dummy;
1575 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1576 }
1577 return 0;
1578}
1579
Dan Gohman0b273252008-11-18 19:49:32 +0000[diff] [blame]1580unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1581 int &FrameIndex) const {
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1582 if (isFrameStoreOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1583 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1584 isFrameOperand(MI, 0, FrameIndex))
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]1585 return MI->getOperand(X86::AddrNumOperands).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1586 return 0;
1587}
1588
1589unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1590 int &FrameIndex) const {
1591 if (isFrameStoreOpcode(MI->getOpcode())) {
1592 unsigned Reg;
1593 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1594 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1595 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1596 const MachineMemOperand *Dummy;
1597 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1598 }
1599 return 0;
1600}
1601
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1602/// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1603/// X86::MOVPC32r.
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1604static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
Jakob Stoklund Olesen3b9a4422012-08-08 00:40:47 +0000[diff] [blame]1605 // Don't waste compile time scanning use-def chains of physregs.
1606 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1607 return false;
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1608 bool isPICBase = false;
1609 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1610 E = MRI.def_end(); I != E; ++I) {
1611 MachineInstr *DefMI = I.getOperand().getParent();
1612 if (DefMI->getOpcode() != X86::MOVPC32r)
1613 return false;
1614 assert(!isPICBase && "More than one PIC base?");
1615 isPICBase = true;
1616 }
1617 return isPICBase;
1618}
Evan Cheng1973a462008-03-31 07:54:19 +0000[diff] [blame]1619
Bill Wendling1e117682008-05-12 20:54:26 +0000[diff] [blame]1620bool
Dan Gohmane919de52009-10-10 00:34:18 +0000[diff] [blame]1621X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1622 AliasAnalysis *AA) const {
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1623 switch (MI->getOpcode()) {
1624 default: break;
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1625 case X86::MOV8rm:
1626 case X86::MOV16rm:
1627 case X86::MOV32rm:
1628 case X86::MOV64rm:
1629 case X86::LD_Fp64m:
1630 case X86::MOVSSrm:
1631 case X86::MOVSDrm:
1632 case X86::MOVAPSrm:
1633 case X86::MOVUPSrm:
1634 case X86::MOVAPDrm:
1635 case X86::MOVDQArm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1636 case X86::MOVDQUrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1637 case X86::VMOVSSrm:
1638 case X86::VMOVSDrm:
1639 case X86::VMOVAPSrm:
1640 case X86::VMOVUPSrm:
1641 case X86::VMOVAPDrm:
1642 case X86::VMOVDQArm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1643 case X86::VMOVDQUrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1644 case X86::VMOVAPSYrm:
1645 case X86::VMOVUPSYrm:
1646 case X86::VMOVAPDYrm:
1647 case X86::VMOVDQAYrm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1648 case X86::VMOVDQUYrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1649 case X86::MMX_MOVD64rm:
1650 case X86::MMX_MOVQ64rm:
1651 case X86::FsVMOVAPSrm:
1652 case X86::FsVMOVAPDrm:
1653 case X86::FsMOVAPSrm:
1654 case X86::FsMOVAPDrm: {
1655 // Loads from constant pools are trivially rematerializable.
1656 if (MI->getOperand(1).isReg() &&
1657 MI->getOperand(2).isImm() &&
1658 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1659 MI->isInvariantLoad(AA)) {
1660 unsigned BaseReg = MI->getOperand(1).getReg();
1661 if (BaseReg == 0 || BaseReg == X86::RIP)
1662 return true;
1663 // Allow re-materialization of PIC load.
1664 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1665 return false;
1666 const MachineFunction &MF = *MI->getParent()->getParent();
1667 const MachineRegisterInfo &MRI = MF.getRegInfo();
1668 return regIsPICBase(BaseReg, MRI);
Evan Cheng94ba37f2008-02-22 09:25:47 +0000[diff] [blame]1669 }
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1670 return false;
1671 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1672
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1673 case X86::LEA32r:
1674 case X86::LEA64r: {
1675 if (MI->getOperand(2).isImm() &&
1676 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1677 !MI->getOperand(4).isReg()) {
1678 // lea fi#, lea GV, etc. are all rematerializable.
1679 if (!MI->getOperand(1).isReg())
1680 return true;
1681 unsigned BaseReg = MI->getOperand(1).getReg();
1682 if (BaseReg == 0)
1683 return true;
1684 // Allow re-materialization of lea PICBase + x.
1685 const MachineFunction &MF = *MI->getParent()->getParent();
1686 const MachineRegisterInfo &MRI = MF.getRegInfo();
1687 return regIsPICBase(BaseReg, MRI);
1688 }
1689 return false;
1690 }
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1691 }
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1692
Dan Gohmane8c1e422007-06-26 00:48:07 +0000[diff] [blame]1693 // All other instructions marked M_REMATERIALIZABLE are always trivially
1694 // rematerializable.
1695 return true;
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1696}
1697
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1698/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1699/// would clobber the EFLAGS condition register. Note the result may be
1700/// conservative. If it cannot definitely determine the safety after visiting
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1701/// a few instructions in each direction it assumes it's not safe.
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1702static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1703 MachineBasicBlock::iterator I) {
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1704 MachineBasicBlock::iterator E = MBB.end();
1705
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1706 // For compile time consideration, if we are not able to determine the
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1707 // safety after visiting 4 instructions in each direction, we will assume
1708 // it's not safe.
1709 MachineBasicBlock::iterator Iter = I;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1710 for (unsigned i = 0; Iter != E && i < 4; ++i) {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1711 bool SeenDef = false;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1712 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1713 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1714 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1715 SeenDef = true;
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1716 if (!MO.isReg())
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1717 continue;
1718 if (MO.getReg() == X86::EFLAGS) {
1719 if (MO.isUse())
1720 return false;
1721 SeenDef = true;
1722 }
1723 }
1724
1725 if (SeenDef)
1726 // This instruction defines EFLAGS, no need to look any further.
1727 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1728 ++Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1729 // Skip over DBG_VALUE.
1730 while (Iter != E && Iter->isDebugValue())
1731 ++Iter;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1732 }
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1733
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1734 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1735 // live in.
1736 if (Iter == E) {
1737 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1738 SE = MBB.succ_end(); SI != SE; ++SI)
1739 if ((*SI)->isLiveIn(X86::EFLAGS))
1740 return false;
1741 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1742 }
1743
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1744 MachineBasicBlock::iterator B = MBB.begin();
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1745 Iter = I;
1746 for (unsigned i = 0; i < 4; ++i) {
1747 // If we make it to the beginning of the block, it's safe to clobber
Sylvestre Ledru91ce36c2012-09-27 10:14:43 +0000[diff] [blame]1748 // EFLAGS iff EFLAGS is not live-in.
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1749 if (Iter == B)
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1750 return !MBB.isLiveIn(X86::EFLAGS);
1751
1752 --Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1753 // Skip over DBG_VALUE.
1754 while (Iter != B && Iter->isDebugValue())
1755 --Iter;
1756
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1757 bool SawKill = false;
1758 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1759 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1760 // A register mask may clobber EFLAGS, but we should still look for a
1761 // live EFLAGS def.
1762 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1763 SawKill = true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1764 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1765 if (MO.isDef()) return MO.isDead();
1766 if (MO.isKill()) SawKill = true;
1767 }
1768 }
1769
1770 if (SawKill)
1771 // This instruction kills EFLAGS and doesn't redefine it, so
1772 // there's no need to look further.
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1773 return true;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1774 }
1775
1776 // Conservative answer.
1777 return false;
1778}
1779
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1780void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1781 MachineBasicBlock::iterator I,
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1782 unsigned DestReg, unsigned SubIdx,
Evan Cheng6ad7da92009-11-14 02:55:43 +0000[diff] [blame]1783 const MachineInstr *Orig,
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1784 const TargetRegisterInfo &TRI) const {
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]1785 // MOV32r0 is implemented with a xor which clobbers condition code.
1786 // Re-materialize it as movri instructions to avoid side effects.
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1787 unsigned Opc = Orig->getOpcode();
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]1788 if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
1789 DebugLoc DL = Orig->getDebugLoc();
1790 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
1791 .addImm(0);
1792 } else {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1793 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1794 MBB.insert(I, MI);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1795 }
Evan Cheng147cb762008-04-16 23:44:44 +0000[diff] [blame]1796
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1797 MachineInstr *NewMI = prior(I);
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1798 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1799}
1800
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1801/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1802/// is not marked dead.
1803static bool hasLiveCondCodeDef(MachineInstr *MI) {
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1804 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1805 MachineOperand &MO = MI->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1806 if (MO.isReg() && MO.isDef() &&
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1807 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1808 return true;
1809 }
1810 }
1811 return false;
1812}
1813
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]1814/// getTruncatedShiftCount - check whether the shift count for a machine operand
1815/// is non-zero.
1816inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
1817 unsigned ShiftAmtOperandIdx) {
1818 // The shift count is six bits with the REX.W prefix and five bits without.
1819 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
1820 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
1821 return Imm & ShiftCountMask;
1822}
1823
1824/// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
1825/// can be represented by a LEA instruction.
1826inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
1827 // Left shift instructions can be transformed into load-effective-address
1828 // instructions if we can encode them appropriately.
1829 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
1830 // The SIB.scale field is two bits wide which means that we can encode any
1831 // shift amount less than 4.
1832 return ShAmt < 4 && ShAmt > 0;
1833}
1834
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]1835bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
1836 unsigned Opc, bool AllowSP,
1837 unsigned &NewSrc, bool &isKill, bool &isUndef,
1838 MachineOperand &ImplicitOp) const {
1839 MachineFunction &MF = *MI->getParent()->getParent();
1840 const TargetRegisterClass *RC;
1841 if (AllowSP) {
1842 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
1843 } else {
1844 RC = Opc != X86::LEA32r ?
1845 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
1846 }
1847 unsigned SrcReg = Src.getReg();
1848
1849 // For both LEA64 and LEA32 the register already has essentially the right
1850 // type (32-bit or 64-bit) we may just need to forbid SP.
1851 if (Opc != X86::LEA64_32r) {
1852 NewSrc = SrcReg;
1853 isKill = Src.isKill();
1854 isUndef = Src.isUndef();
1855
1856 if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
1857 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
1858 return false;
1859
1860 return true;
1861 }
1862
1863 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
1864 // another we need to add 64-bit registers to the final MI.
1865 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
1866 ImplicitOp = Src;
1867 ImplicitOp.setImplicit();
1868
1869 NewSrc = getX86SubSuperRegister(Src.getReg(), MVT::i64);
1870 MachineBasicBlock::LivenessQueryResult LQR =
1871 MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
1872
1873 switch (LQR) {
1874 case MachineBasicBlock::LQR_Unknown:
1875 // We can't give sane liveness flags to the instruction, abandon LEA
1876 // formation.
1877 return false;
1878 case MachineBasicBlock::LQR_Live:
1879 isKill = MI->killsRegister(SrcReg);
1880 isUndef = false;
1881 break;
1882 default:
1883 // The physreg itself is dead, so we have to use it as an <undef>.
1884 isKill = false;
1885 isUndef = true;
1886 break;
1887 }
1888 } else {
1889 // Virtual register of the wrong class, we have to create a temporary 64-bit
1890 // vreg to feed into the LEA.
1891 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
1892 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1893 get(TargetOpcode::COPY))
1894 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
1895 .addOperand(Src);
1896
1897 // Which is obviously going to be dead after we're done with it.
1898 isKill = true;
1899 isUndef = false;
1900 }
1901
1902 // We've set all the parameters without issue.
1903 return true;
1904}
1905
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1906/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1907/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1908/// to a 32-bit superregister and then truncating back down to a 16-bit
1909/// subregister.
1910MachineInstr *
1911X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1912 MachineFunction::iterator &MFI,
1913 MachineBasicBlock::iterator &MBBI,
1914 LiveVariables *LV) const {
1915 MachineInstr *MI = MBBI;
1916 unsigned Dest = MI->getOperand(0).getReg();
1917 unsigned Src = MI->getOperand(1).getReg();
1918 bool isDead = MI->getOperand(0).isDead();
1919 bool isKill = MI->getOperand(1).isKill();
1920
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1921 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1922 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]1923 unsigned Opc, leaInReg;
1924 if (TM.getSubtarget<X86Subtarget>().is64Bit()) {
1925 Opc = X86::LEA64_32r;
1926 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1927 } else {
1928 Opc = X86::LEA32r;
1929 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1930 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1931
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1932 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1933 // well be shifting and then extracting the lower 16-bits.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1934 // This has the potential to cause partial register stall. e.g.
Evan Cheng3974c8d2009-12-12 18:55:26 +0000[diff] [blame]1935 // movw (%rbp,%rcx,2), %dx
1936 // leal -65(%rdx), %esi
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1937 // But testing has shown this *does* help performance in 64-bit mode (at
1938 // least on modern x86 machines).
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1939 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1940 MachineInstr *InsMI =
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1941 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1942 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1943 .addReg(Src, getKillRegState(isKill));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1944
1945 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1946 get(Opc), leaOutReg);
1947 switch (MIOpc) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1948 default: llvm_unreachable("Unreachable!");
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1949 case X86::SHL16ri: {
1950 unsigned ShAmt = MI->getOperand(2).getImm();
1951 MIB.addReg(0).addImm(1 << ShAmt)
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1952 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1953 break;
1954 }
1955 case X86::INC16r:
1956 case X86::INC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1957 addRegOffset(MIB, leaInReg, true, 1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1958 break;
1959 case X86::DEC16r:
1960 case X86::DEC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1961 addRegOffset(MIB, leaInReg, true, -1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1962 break;
1963 case X86::ADD16ri:
1964 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]1965 case X86::ADD16ri_DB:
1966 case X86::ADD16ri8_DB:
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1967 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1968 break;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1969 case X86::ADD16rr:
1970 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1971 unsigned Src2 = MI->getOperand(2).getReg();
1972 bool isKill2 = MI->getOperand(2).isKill();
1973 unsigned leaInReg2 = 0;
1974 MachineInstr *InsMI2 = 0;
1975 if (Src == Src2) {
1976 // ADD16rr %reg1028<kill>, %reg1028
1977 // just a single insert_subreg.
1978 addRegReg(MIB, leaInReg, true, leaInReg, false);
1979 } else {
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]1980 if (TM.getSubtarget<X86Subtarget>().is64Bit())
1981 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1982 else
1983 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1984 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1985 // well be shifting and then extracting the lower 16-bits.
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1986 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1987 InsMI2 =
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1988 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1989 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1990 .addReg(Src2, getKillRegState(isKill2));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1991 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1992 }
1993 if (LV && isKill2 && InsMI2)
1994 LV->replaceKillInstruction(Src2, MI, InsMI2);
1995 break;
1996 }
1997 }
1998
1999 MachineInstr *NewMI = MIB;
2000 MachineInstr *ExtMI =
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]2001 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2002 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]2003 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2004
2005 if (LV) {
2006 // Update live variables
2007 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
2008 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
2009 if (isKill)
2010 LV->replaceKillInstruction(Src, MI, InsMI);
2011 if (isDead)
2012 LV->replaceKillInstruction(Dest, MI, ExtMI);
2013 }
2014
2015 return ExtMI;
2016}
2017
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2018/// convertToThreeAddress - This method must be implemented by targets that
2019/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
2020/// may be able to convert a two-address instruction into a true
2021/// three-address instruction on demand. This allows the X86 target (for
2022/// example) to convert ADD and SHL instructions into LEA instructions if they
2023/// would require register copies due to two-addressness.
2024///
2025/// This method returns a null pointer if the transformation cannot be
2026/// performed, otherwise it returns the new instruction.
2027///
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]2028MachineInstr *
2029X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
2030 MachineBasicBlock::iterator &MBBI,
Owen Anderson30cc0282008-07-02 23:41:07 +0000[diff] [blame]2031 LiveVariables *LV) const {
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]2032 MachineInstr *MI = MBBI;
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]2033
2034 // The following opcodes also sets the condition code register(s). Only
2035 // convert them to equivalent lea if the condition code register def's
2036 // are dead!
2037 if (hasLiveCondCodeDef(MI))
2038 return 0;
2039
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]2040 MachineFunction &MF = *MI->getParent()->getParent();
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2041 // All instructions input are two-addr instructions. Get the known operands.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2042 const MachineOperand &Dest = MI->getOperand(0);
2043 const MachineOperand &Src = MI->getOperand(1);
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2044
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]2045 MachineInstr *NewMI = NULL;
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]2046 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2047 // we have better subtarget support, enable the 16-bit LEA generation here.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2048 // 16-bit LEA is also slow on Core2.
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]2049 bool DisableLEA16 = true;
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2050 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]2051
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2052 unsigned MIOpc = MI->getOpcode();
2053 switch (MIOpc) {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]2054 case X86::SHUFPSrri: {
2055 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2056 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2057
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]2058 unsigned B = MI->getOperand(1).getReg();
2059 unsigned C = MI->getOperand(2).getReg();
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2060 if (B != C) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2061 unsigned M = MI->getOperand(3).getImm();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2062 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2063 .addOperand(Dest).addOperand(Src).addImm(M);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2064 break;
2065 }
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]2066 case X86::SHUFPDrri: {
2067 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
2068 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
2069
2070 unsigned B = MI->getOperand(1).getReg();
2071 unsigned C = MI->getOperand(2).getReg();
2072 if (B != C) return 0;
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]2073 unsigned M = MI->getOperand(3).getImm();
2074
2075 // Convert to PSHUFD mask.
2076 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
2077
2078 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2079 .addOperand(Dest).addOperand(Src).addImm(M);
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]2080 break;
2081 }
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]2082 case X86::SHL64ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]2083 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]2084 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2085 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2086
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2087 // LEA can't handle RSP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2088 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2089 !MF.getRegInfo().constrainRegClass(Src.getReg(),
2090 &X86::GR64_NOSPRegClass))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2091 return 0;
2092
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2093 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2094 .addOperand(Dest)
2095 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]2096 break;
2097 }
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2098 case X86::SHL32ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]2099 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]2100 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2101 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2102
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2103 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2104
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2105 // LEA can't handle ESP.
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2106 bool isKill, isUndef;
2107 unsigned SrcReg;
2108 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2109 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2110 SrcReg, isKill, isUndef, ImplicitOp))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2111 return 0;
2112
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2113 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2114 .addOperand(Dest)
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2115 .addReg(0).addImm(1 << ShAmt)
2116 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2117 .addImm(0).addReg(0);
2118 if (ImplicitOp.getReg() != 0)
2119 MIB.addOperand(ImplicitOp);
2120 NewMI = MIB;
2121
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2122 break;
2123 }
2124 case X86::SHL16ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]2125 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]2126 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2127 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2128
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2129 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2130 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2131 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2132 .addOperand(Dest)
2133 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2134 break;
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]2135 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2136 default: {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]2137
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2138 switch (MIOpc) {
2139 default: return 0;
2140 case X86::INC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]2141 case X86::INC32r:
2142 case X86::INC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2143 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]2144 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2145 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2146 bool isKill, isUndef;
2147 unsigned SrcReg;
2148 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2149 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2150 SrcReg, isKill, isUndef, ImplicitOp))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2151 return 0;
2152
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2153 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2154 .addOperand(Dest)
2155 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef));
2156 if (ImplicitOp.getReg() != 0)
2157 MIB.addOperand(ImplicitOp);
2158
2159 NewMI = addOffset(MIB, 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2160 break;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2161 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2162 case X86::INC16r:
2163 case X86::INC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2164 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2165 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2166 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2167 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2168 .addOperand(Dest).addOperand(Src), 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2169 break;
2170 case X86::DEC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]2171 case X86::DEC32r:
2172 case X86::DEC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2173 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]2174 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2175 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2176
2177 bool isKill, isUndef;
2178 unsigned SrcReg;
2179 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2180 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2181 SrcReg, isKill, isUndef, ImplicitOp))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2182 return 0;
2183
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2184 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2185 .addOperand(Dest)
2186 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2187 if (ImplicitOp.getReg() != 0)
2188 MIB.addOperand(ImplicitOp);
2189
2190 NewMI = addOffset(MIB, -1);
2191
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2192 break;
2193 }
2194 case X86::DEC16r:
2195 case X86::DEC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2196 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2197 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2198 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2199 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2200 .addOperand(Dest).addOperand(Src), -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2201 break;
2202 case X86::ADD64rr:
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2203 case X86::ADD64rr_DB:
2204 case X86::ADD32rr:
2205 case X86::ADD32rr_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2206 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2207 unsigned Opc;
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2208 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2209 Opc = X86::LEA64r;
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2210 else
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2211 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2212
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2213 bool isKill, isUndef;
2214 unsigned SrcReg;
2215 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2216 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2217 SrcReg, isKill, isUndef, ImplicitOp))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2218 return 0;
2219
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2220 const MachineOperand &Src2 = MI->getOperand(2);
2221 bool isKill2, isUndef2;
2222 unsigned SrcReg2;
2223 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2224 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2225 SrcReg2, isKill2, isUndef2, ImplicitOp2))
2226 return 0;
2227
2228 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2229 .addOperand(Dest);
2230 if (ImplicitOp.getReg() != 0)
2231 MIB.addOperand(ImplicitOp);
2232 if (ImplicitOp2.getReg() != 0)
2233 MIB.addOperand(ImplicitOp2);
2234
2235 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
Nadav Rotem4968e452012-07-16 10:52:25 +0000[diff] [blame]2236
2237 // Preserve undefness of the operands.
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2238 NewMI->getOperand(1).setIsUndef(isUndef);
2239 NewMI->getOperand(3).setIsUndef(isUndef2);
Nadav Rotem4968e452012-07-16 10:52:25 +0000[diff] [blame]2240
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2241 if (LV && Src2.isKill())
2242 LV->replaceKillInstruction(SrcReg2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2243 break;
2244 }
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2245 case X86::ADD16rr:
2246 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2247 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2248 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2249 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2250 unsigned Src2 = MI->getOperand(2).getReg();
2251 bool isKill2 = MI->getOperand(2).isKill();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2252 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2253 .addOperand(Dest),
2254 Src.getReg(), Src.isKill(), Src2, isKill2);
2255
2256 // Preserve undefness of the operands.
2257 bool isUndef = MI->getOperand(1).isUndef();
2258 bool isUndef2 = MI->getOperand(2).isUndef();
2259 NewMI->getOperand(1).setIsUndef(isUndef);
2260 NewMI->getOperand(3).setIsUndef(isUndef2);
2261
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2262 if (LV && isKill2)
2263 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2264 break;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2265 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2266 case X86::ADD64ri32:
2267 case X86::ADD64ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2268 case X86::ADD64ri32_DB:
2269 case X86::ADD64ri8_DB:
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2270 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2271 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2272 .addOperand(Dest).addOperand(Src),
2273 MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2274 break;
2275 case X86::ADD32ri:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2276 case X86::ADD32ri8:
2277 case X86::ADD32ri_DB:
2278 case X86::ADD32ri8_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2279 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2280 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2281
2282 bool isKill, isUndef;
2283 unsigned SrcReg;
2284 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2285 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2286 SrcReg, isKill, isUndef, ImplicitOp))
2287 return 0;
2288
2289 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2290 .addOperand(Dest)
2291 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2292 if (ImplicitOp.getReg() != 0)
2293 MIB.addOperand(ImplicitOp);
2294
2295 NewMI = addOffset(MIB, MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2296 break;
2297 }
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2298 case X86::ADD16ri:
2299 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2300 case X86::ADD16ri_DB:
2301 case X86::ADD16ri8_DB:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2302 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2303 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2304 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2305 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2306 .addOperand(Dest).addOperand(Src),
2307 MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2308 break;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2309 }
2310 }
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2311 }
2312
Evan Cheng1bc1cae2008-02-07 08:29:53 +0000[diff] [blame]2313 if (!NewMI) return 0;
2314
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2315 if (LV) { // Update live variables
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2316 if (Src.isKill())
2317 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2318 if (Dest.isDead())
2319 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2320 }
2321
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2322 MFI->insert(MBBI, NewMI); // Insert the new inst
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]2323 return NewMI;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2324}
2325
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2326/// commuteInstruction - We have a few instructions that must be hacked on to
2327/// commute them.
2328///
Evan Cheng03553bb2008-06-16 07:33:11 +0000[diff] [blame]2329MachineInstr *
2330X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2331 switch (MI->getOpcode()) {
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2332 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2333 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2334 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2335 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2336 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2337 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2338 unsigned Opc;
2339 unsigned Size;
2340 switch (MI->getOpcode()) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2341 default: llvm_unreachable("Unreachable!");
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2342 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2343 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2344 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2345 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2346 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2347 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2348 }
Chris Lattner5c463782007-12-30 20:49:49 +0000[diff] [blame]2349 unsigned Amt = MI->getOperand(3).getImm();
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2350 if (NewMI) {
2351 MachineFunction &MF = *MI->getParent()->getParent();
2352 MI = MF.CloneMachineInstr(MI);
2353 NewMI = false;
Evan Cheng244183e2008-02-13 02:46:49 +0000[diff] [blame]2354 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2355 MI->setDesc(get(Opc));
2356 MI->getOperand(3).setImm(Size-Amt);
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]2357 return TargetInstrInfo::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2358 }
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2359 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2360 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2361 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2362 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2363 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2364 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2365 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2366 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2367 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2368 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2369 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2370 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2371 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2372 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2373 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2374 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2375 unsigned Opc;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2376 switch (MI->getOpcode()) {
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2377 default: llvm_unreachable("Unreachable!");
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2378 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2379 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2380 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2381 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2382 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2383 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2384 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2385 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2386 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2387 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2388 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2389 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
Chris Lattner1a1c6002010-10-05 23:00:14 +0000[diff] [blame]2390 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2391 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2392 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2393 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2394 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2395 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2396 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2397 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2398 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2399 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2400 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2401 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2402 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2403 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2404 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2405 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2406 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2407 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2408 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2409 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2410 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2411 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2412 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2413 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2414 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2415 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2416 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2417 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2418 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2419 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2420 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2421 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2422 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2423 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2424 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2425 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2426 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2427 if (NewMI) {
2428 MachineFunction &MF = *MI->getParent()->getParent();
2429 MI = MF.CloneMachineInstr(MI);
2430 NewMI = false;
2431 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]2432 MI->setDesc(get(Opc));
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2433 // Fallthrough intended.
2434 }
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2435 default:
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]2436 return TargetInstrInfo::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2437 }
2438}
2439
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2440static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2441 switch (BrOpc) {
2442 default: return X86::COND_INVALID;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2443 case X86::JE_4: return X86::COND_E;
2444 case X86::JNE_4: return X86::COND_NE;
2445 case X86::JL_4: return X86::COND_L;
2446 case X86::JLE_4: return X86::COND_LE;
2447 case X86::JG_4: return X86::COND_G;
2448 case X86::JGE_4: return X86::COND_GE;
2449 case X86::JB_4: return X86::COND_B;
2450 case X86::JBE_4: return X86::COND_BE;
2451 case X86::JA_4: return X86::COND_A;
2452 case X86::JAE_4: return X86::COND_AE;
2453 case X86::JS_4: return X86::COND_S;
2454 case X86::JNS_4: return X86::COND_NS;
2455 case X86::JP_4: return X86::COND_P;
2456 case X86::JNP_4: return X86::COND_NP;
2457 case X86::JO_4: return X86::COND_O;
2458 case X86::JNO_4: return X86::COND_NO;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2459 }
2460}
2461
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2462/// getCondFromSETOpc - return condition code of a SET opcode.
2463static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2464 switch (Opc) {
2465 default: return X86::COND_INVALID;
2466 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2467 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2468 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2469 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2470 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2471 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2472 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2473 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2474 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2475 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2476 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2477 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2478 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2479 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2480 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2481 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2482 }
2483}
2484
2485/// getCondFromCmovOpc - return condition code of a CMov opcode.
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]2486X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2487 switch (Opc) {
2488 default: return X86::COND_INVALID;
2489 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2490 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2491 return X86::COND_A;
2492 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2493 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2494 return X86::COND_AE;
2495 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2496 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2497 return X86::COND_B;
2498 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2499 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2500 return X86::COND_BE;
2501 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2502 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2503 return X86::COND_E;
2504 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2505 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2506 return X86::COND_G;
2507 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2508 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2509 return X86::COND_GE;
2510 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2511 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2512 return X86::COND_L;
2513 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2514 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2515 return X86::COND_LE;
2516 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2517 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2518 return X86::COND_NE;
2519 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2520 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2521 return X86::COND_NO;
2522 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2523 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2524 return X86::COND_NP;
2525 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2526 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2527 return X86::COND_NS;
2528 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2529 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2530 return X86::COND_O;
2531 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2532 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2533 return X86::COND_P;
2534 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2535 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2536 return X86::COND_S;
2537 }
2538}
2539
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2540unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2541 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2542 default: llvm_unreachable("Illegal condition code!");
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2543 case X86::COND_E: return X86::JE_4;
2544 case X86::COND_NE: return X86::JNE_4;
2545 case X86::COND_L: return X86::JL_4;
2546 case X86::COND_LE: return X86::JLE_4;
2547 case X86::COND_G: return X86::JG_4;
2548 case X86::COND_GE: return X86::JGE_4;
2549 case X86::COND_B: return X86::JB_4;
2550 case X86::COND_BE: return X86::JBE_4;
2551 case X86::COND_A: return X86::JA_4;
2552 case X86::COND_AE: return X86::JAE_4;
2553 case X86::COND_S: return X86::JS_4;
2554 case X86::COND_NS: return X86::JNS_4;
2555 case X86::COND_P: return X86::JP_4;
2556 case X86::COND_NP: return X86::JNP_4;
2557 case X86::COND_O: return X86::JO_4;
2558 case X86::COND_NO: return X86::JNO_4;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2559 }
2560}
2561
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2562/// GetOppositeBranchCondition - Return the inverse of the specified condition,
2563/// e.g. turning COND_E to COND_NE.
2564X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2565 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2566 default: llvm_unreachable("Illegal condition code!");
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2567 case X86::COND_E: return X86::COND_NE;
2568 case X86::COND_NE: return X86::COND_E;
2569 case X86::COND_L: return X86::COND_GE;
2570 case X86::COND_LE: return X86::COND_G;
2571 case X86::COND_G: return X86::COND_LE;
2572 case X86::COND_GE: return X86::COND_L;
2573 case X86::COND_B: return X86::COND_AE;
2574 case X86::COND_BE: return X86::COND_A;
2575 case X86::COND_A: return X86::COND_BE;
2576 case X86::COND_AE: return X86::COND_B;
2577 case X86::COND_S: return X86::COND_NS;
2578 case X86::COND_NS: return X86::COND_S;
2579 case X86::COND_P: return X86::COND_NP;
2580 case X86::COND_NP: return X86::COND_P;
2581 case X86::COND_O: return X86::COND_NO;
2582 case X86::COND_NO: return X86::COND_O;
2583 }
2584}
2585
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2586/// getSwappedCondition - assume the flags are set by MI(a,b), return
2587/// the condition code if we modify the instructions such that flags are
2588/// set by MI(b,a).
Benjamin Kramerabbfe692012-07-13 13:25:15 +0000[diff] [blame]2589static X86::CondCode getSwappedCondition(X86::CondCode CC) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2590 switch (CC) {
2591 default: return X86::COND_INVALID;
2592 case X86::COND_E: return X86::COND_E;
2593 case X86::COND_NE: return X86::COND_NE;
2594 case X86::COND_L: return X86::COND_G;
2595 case X86::COND_LE: return X86::COND_GE;
2596 case X86::COND_G: return X86::COND_L;
2597 case X86::COND_GE: return X86::COND_LE;
2598 case X86::COND_B: return X86::COND_A;
2599 case X86::COND_BE: return X86::COND_AE;
2600 case X86::COND_A: return X86::COND_B;
2601 case X86::COND_AE: return X86::COND_BE;
2602 }
2603}
2604
2605/// getSETFromCond - Return a set opcode for the given condition and
2606/// whether it has memory operand.
2607static unsigned getSETFromCond(X86::CondCode CC,
2608 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2609 static const uint16_t Opc[16][2] = {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2610 { X86::SETAr, X86::SETAm },
2611 { X86::SETAEr, X86::SETAEm },
2612 { X86::SETBr, X86::SETBm },
2613 { X86::SETBEr, X86::SETBEm },
2614 { X86::SETEr, X86::SETEm },
2615 { X86::SETGr, X86::SETGm },
2616 { X86::SETGEr, X86::SETGEm },
2617 { X86::SETLr, X86::SETLm },
2618 { X86::SETLEr, X86::SETLEm },
2619 { X86::SETNEr, X86::SETNEm },
2620 { X86::SETNOr, X86::SETNOm },
2621 { X86::SETNPr, X86::SETNPm },
2622 { X86::SETNSr, X86::SETNSm },
2623 { X86::SETOr, X86::SETOm },
2624 { X86::SETPr, X86::SETPm },
2625 { X86::SETSr, X86::SETSm }
2626 };
2627
2628 assert(CC < 16 && "Can only handle standard cond codes");
2629 return Opc[CC][HasMemoryOperand ? 1 : 0];
2630}
2631
2632/// getCMovFromCond - Return a cmov opcode for the given condition,
2633/// register size in bytes, and operand type.
2634static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2635 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2636 static const uint16_t Opc[32][3] = {
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2637 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2638 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2639 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2640 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2641 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2642 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2643 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2644 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2645 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2646 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2647 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2648 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2649 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2650 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2651 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2652 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2653 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2654 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2655 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2656 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2657 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2658 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2659 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2660 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2661 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2662 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2663 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2664 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2665 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2666 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2667 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2668 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2669 };
2670
2671 assert(CC < 16 && "Can only handle standard cond codes");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2672 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2673 switch(RegBytes) {
2674 default: llvm_unreachable("Illegal register size!");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2675 case 2: return Opc[Idx][0];
2676 case 4: return Opc[Idx][1];
2677 case 8: return Opc[Idx][2];
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2678 }
2679}
2680
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2681bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2682 if (!MI->isTerminator()) return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2683
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2684 // Conditional branch is a special case.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2685 if (MI->isBranch() && !MI->isBarrier())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2686 return true;
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2687 if (!MI->isPredicable())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2688 return true;
2689 return !isPredicated(MI);
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2690}
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2691
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2692bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2693 MachineBasicBlock *&TBB,
2694 MachineBasicBlock *&FBB,
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2695 SmallVectorImpl<MachineOperand> &Cond,
2696 bool AllowModify) const {
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2697 // Start from the bottom of the block and work up, examining the
2698 // terminator instructions.
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2699 MachineBasicBlock::iterator I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2700 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2701 while (I != MBB.begin()) {
2702 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2703 if (I->isDebugValue())
2704 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2705
2706 // Working from the bottom, when we see a non-terminator instruction, we're
2707 // done.
Jakob Stoklund Olesenc30b4dd2010-07-16 17:41:44 +0000[diff] [blame]2708 if (!isUnpredicatedTerminator(I))
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2709 break;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2710
2711 // A terminator that isn't a branch can't easily be handled by this
2712 // analysis.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2713 if (!I->isBranch())
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2714 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2715
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2716 // Handle unconditional branches.
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2717 if (I->getOpcode() == X86::JMP_4) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2718 UnCondBrIter = I;
2719
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2720 if (!AllowModify) {
2721 TBB = I->getOperand(0).getMBB();
Evan Cheng2fa28112009-05-08 06:34:09 +0000[diff] [blame]2722 continue;
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2723 }
2724
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2725 // If the block has any instructions after a JMP, delete them.
Chris Lattnera48f44d2009-12-03 00:50:42 +0000[diff] [blame]2726 while (llvm::next(I) != MBB.end())
2727 llvm::next(I)->eraseFromParent();
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2728
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2729 Cond.clear();
2730 FBB = 0;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2731
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2732 // Delete the JMP if it's equivalent to a fall-through.
2733 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2734 TBB = 0;
2735 I->eraseFromParent();
2736 I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2737 UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2738 continue;
2739 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2740
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2741 // TBB is used to indicate the unconditional destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2742 TBB = I->getOperand(0).getMBB();
2743 continue;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2744 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2745
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2746 // Handle conditional branches.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2747 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2748 if (BranchCode == X86::COND_INVALID)
2749 return true; // Can't handle indirect branch.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2750
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2751 // Working from the bottom, handle the first conditional branch.
2752 if (Cond.empty()) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2753 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2754 if (AllowModify && UnCondBrIter != MBB.end() &&
2755 MBB.isLayoutSuccessor(TargetBB)) {
2756 // If we can modify the code and it ends in something like:
2757 //
2758 // jCC L1
2759 // jmp L2
2760 // L1:
2761 // ...
2762 // L2:
2763 //
2764 // Then we can change this to:
2765 //
2766 // jnCC L2
2767 // L1:
2768 // ...
2769 // L2:
2770 //
2771 // Which is a bit more efficient.
2772 // We conditionally jump to the fall-through block.
2773 BranchCode = GetOppositeBranchCondition(BranchCode);
2774 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2775 MachineBasicBlock::iterator OldInst = I;
2776
2777 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2778 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2779 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2780 .addMBB(TargetBB);
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2781
2782 OldInst->eraseFromParent();
2783 UnCondBrIter->eraseFromParent();
2784
2785 // Restart the analysis.
2786 UnCondBrIter = MBB.end();
2787 I = MBB.end();
2788 continue;
2789 }
2790
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2791 FBB = TBB;
2792 TBB = I->getOperand(0).getMBB();
2793 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2794 continue;
2795 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2796
2797 // Handle subsequent conditional branches. Only handle the case where all
2798 // conditional branches branch to the same destination and their condition
2799 // opcodes fit one of the special multi-branch idioms.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2800 assert(Cond.size() == 1);
2801 assert(TBB);
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2802
2803 // Only handle the case where all conditional branches branch to the same
2804 // destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2805 if (TBB != I->getOperand(0).getMBB())
2806 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2807
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2808 // If the conditions are the same, we can leave them alone.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2809 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2810 if (OldBranchCode == BranchCode)
2811 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2812
2813 // If they differ, see if they fit one of the known patterns. Theoretically,
2814 // we could handle more patterns here, but we shouldn't expect to see them
2815 // if instruction selection has done a reasonable job.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2816 if ((OldBranchCode == X86::COND_NP &&
2817 BranchCode == X86::COND_E) ||
2818 (OldBranchCode == X86::COND_E &&
2819 BranchCode == X86::COND_NP))
2820 BranchCode = X86::COND_NP_OR_E;
2821 else if ((OldBranchCode == X86::COND_P &&
2822 BranchCode == X86::COND_NE) ||
2823 (OldBranchCode == X86::COND_NE &&
2824 BranchCode == X86::COND_P))
2825 BranchCode = X86::COND_NE_OR_P;
2826 else
2827 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2828
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2829 // Update the MachineOperand.
2830 Cond[0].setImm(BranchCode);
Chris Lattner74436002006-10-30 22:27:23 +0000[diff] [blame]2831 }
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2832
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2833 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2834}
2835
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2836unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2837 MachineBasicBlock::iterator I = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2838 unsigned Count = 0;
2839
2840 while (I != MBB.begin()) {
2841 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2842 if (I->isDebugValue())
2843 continue;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2844 if (I->getOpcode() != X86::JMP_4 &&
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2845 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2846 break;
2847 // Remove the branch.
2848 I->eraseFromParent();
2849 I = MBB.end();
2850 ++Count;
2851 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2852
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2853 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2854}
2855
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2856unsigned
2857X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2858 MachineBasicBlock *FBB,
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2859 const SmallVectorImpl<MachineOperand> &Cond,
2860 DebugLoc DL) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2861 // Shouldn't be a fall through.
2862 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
Chris Lattner6fca75e2006-10-21 05:34:23 +0000[diff] [blame]2863 assert((Cond.size() == 1 || Cond.size() == 0) &&
2864 "X86 branch conditions have one component!");
2865
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2866 if (Cond.empty()) {
2867 // Unconditional branch?
2868 assert(!FBB && "Unconditional branch with multiple successors!");
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2869 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2870 return 1;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2871 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2872
2873 // Conditional branch.
2874 unsigned Count = 0;
2875 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2876 switch (CC) {
2877 case X86::COND_NP_OR_E:
2878 // Synthesize NP_OR_E with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2879 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2880 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2881 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2882 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2883 break;
2884 case X86::COND_NE_OR_P:
2885 // Synthesize NE_OR_P with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2886 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2887 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2888 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2889 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2890 break;
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2891 default: {
2892 unsigned Opc = GetCondBranchFromCond(CC);
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2893 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2894 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2895 }
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2896 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2897 if (FBB) {
2898 // Two-way Conditional branch. Insert the second branch.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2899 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2900 ++Count;
2901 }
2902 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2903}
2904
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2905bool X86InstrInfo::
2906canInsertSelect(const MachineBasicBlock &MBB,
2907 const SmallVectorImpl<MachineOperand> &Cond,
2908 unsigned TrueReg, unsigned FalseReg,
2909 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2910 // Not all subtargets have cmov instructions.
2911 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2912 return false;
2913 if (Cond.size() != 1)
2914 return false;
2915 // We cannot do the composite conditions, at least not in SSA form.
2916 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2917 return false;
2918
2919 // Check register classes.
2920 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2921 const TargetRegisterClass *RC =
2922 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2923 if (!RC)
2924 return false;
2925
2926 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2927 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2928 X86::GR32RegClass.hasSubClassEq(RC) ||
2929 X86::GR64RegClass.hasSubClassEq(RC)) {
2930 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2931 // Bridge. Probably Ivy Bridge as well.
2932 CondCycles = 2;
2933 TrueCycles = 2;
2934 FalseCycles = 2;
2935 return true;
2936 }
2937
2938 // Can't do vectors.
2939 return false;
2940}
2941
2942void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2943 MachineBasicBlock::iterator I, DebugLoc DL,
2944 unsigned DstReg,
2945 const SmallVectorImpl<MachineOperand> &Cond,
2946 unsigned TrueReg, unsigned FalseReg) const {
2947 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2948 assert(Cond.size() == 1 && "Invalid Cond array");
2949 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2950 MRI.getRegClass(DstReg)->getSize(),
2951 false/*HasMemoryOperand*/);
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2952 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2953}
2954
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2955/// isHReg - Test if the given register is a physical h register.
2956static bool isHReg(unsigned Reg) {
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2957 return X86::GR8_ABCD_HRegClass.contains(Reg);
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2958}
2959
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2960// Try and copy between VR128/VR64 and GR64 registers.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2961static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]2962 const X86Subtarget& Subtarget) {
2963
2964
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2965 // SrcReg(VR128) -> DestReg(GR64)
2966 // SrcReg(VR64) -> DestReg(GR64)
2967 // SrcReg(GR64) -> DestReg(VR128)
2968 // SrcReg(GR64) -> DestReg(VR64)
2969
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]2970 bool HasAVX = Subtarget.hasAVX();
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2971 bool HasAVX512 = Subtarget.hasAVX512();
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2972 if (X86::GR64RegClass.contains(DestReg)) {
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2973 if (X86::VR128XRegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2974 // Copy from a VR128 register to a GR64 register.
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2975 return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
2976 X86::MOVPQIto64rr);
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2977 if (X86::VR64RegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2978 // Copy from a VR64 register to a GR64 register.
2979 return X86::MOVSDto64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2980 } else if (X86::GR64RegClass.contains(SrcReg)) {
2981 // Copy from a GR64 register to a VR128 register.
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2982 if (X86::VR128XRegClass.contains(DestReg))
2983 return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
2984 X86::MOV64toPQIrr);
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2985 // Copy from a GR64 register to a VR64 register.
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2986 if (X86::VR64RegClass.contains(DestReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2987 return X86::MOV64toSDrr;
2988 }
2989
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2990 // SrcReg(FR32) -> DestReg(GR32)
2991 // SrcReg(GR32) -> DestReg(FR32)
2992
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2993 if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2994 // Copy from a FR32 register to a GR32 register.
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2995 return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2996
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2997 if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2998 // Copy from a GR32 register to a FR32 register.
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2999 return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]3000 return 0;
3001}
3002
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]3003static
3004unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg) {
3005 if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
3006 X86::VR256XRegClass.contains(DestReg, SrcReg) ||
3007 X86::VR512RegClass.contains(DestReg, SrcReg)) {
3008 DestReg = get512BitSuperRegister(DestReg);
3009 SrcReg = get512BitSuperRegister(SrcReg);
3010 return X86::VMOVAPSZrr;
3011 }
3012 if ((X86::VK8RegClass.contains(DestReg) ||
3013 X86::VK16RegClass.contains(DestReg)) &&
3014 (X86::VK8RegClass.contains(SrcReg) ||
3015 X86::VK16RegClass.contains(SrcReg)))
3016 return X86::KMOVWkk;
3017 return 0;
3018}
3019
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3020void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3021 MachineBasicBlock::iterator MI, DebugLoc DL,
3022 unsigned DestReg, unsigned SrcReg,
3023 bool KillSrc) const {
3024 // First deal with the normal symmetric copies.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3025 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]3026 bool HasAVX512 = TM.getSubtarget<X86Subtarget>().hasAVX512();
3027 unsigned Opc = 0;
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3028 if (X86::GR64RegClass.contains(DestReg, SrcReg))
3029 Opc = X86::MOV64rr;
3030 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3031 Opc = X86::MOV32rr;
3032 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3033 Opc = X86::MOV16rr;
3034 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3035 // Copying to or from a physical H register on x86-64 requires a NOREX
3036 // move. Otherwise use a normal move.
3037 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]3038 TM.getSubtarget<X86Subtarget>().is64Bit()) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3039 Opc = X86::MOV8rr_NOREX;
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]3040 // Both operands must be encodable without an REX prefix.
3041 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3042 "8-bit H register can not be copied outside GR8_NOREX");
3043 } else
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3044 Opc = X86::MOV8rr;
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]3045 }
3046 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3047 Opc = X86::MMX_MOVQ64rr;
3048 else if (HasAVX512)
3049 Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg);
3050 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3051 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]3052 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3053 Opc = X86::VMOVAPSYrr;
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]3054 if (!Opc)
3055 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, TM.getSubtarget<X86Subtarget>());
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3056
3057 if (Opc) {
3058 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3059 .addReg(SrcReg, getKillRegState(KillSrc));
3060 return;
3061 }
3062
3063 // Moving EFLAGS to / from another register requires a push and a pop.
Nadav Rotemd5aae982012-12-21 23:48:49 +0000[diff] [blame]3064 // Notice that we have to adjust the stack if we don't want to clobber the
3065 // first frame index. See X86FrameLowering.cpp - colobbersTheStack.
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3066 if (SrcReg == X86::EFLAGS) {
3067 if (X86::GR64RegClass.contains(DestReg)) {
3068 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
3069 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
3070 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]3071 }
3072 if (X86::GR32RegClass.contains(DestReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3073 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
3074 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
3075 return;
3076 }
3077 }
3078 if (DestReg == X86::EFLAGS) {
3079 if (X86::GR64RegClass.contains(SrcReg)) {
3080 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
3081 .addReg(SrcReg, getKillRegState(KillSrc));
3082 BuildMI(MBB, MI, DL, get(X86::POPF64));
3083 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]3084 }
3085 if (X86::GR32RegClass.contains(SrcReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3086 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
3087 .addReg(SrcReg, getKillRegState(KillSrc));
3088 BuildMI(MBB, MI, DL, get(X86::POPF32));
3089 return;
3090 }
3091 }
3092
3093 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
3094 << " to " << RI.getName(DestReg) << '\n');
3095 llvm_unreachable("Cannot emit physreg copy instruction");
3096}
3097
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3098static unsigned getLoadStoreRegOpcode(unsigned Reg,
3099 const TargetRegisterClass *RC,
3100 bool isStackAligned,
3101 const TargetMachine &TM,
3102 bool load) {
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3103 if (TM.getSubtarget<X86Subtarget>().hasAVX512()) {
3104 if (X86::VK8RegClass.hasSubClassEq(RC) ||
3105 X86::VK16RegClass.hasSubClassEq(RC))
3106 return load ? X86::KMOVWkm : X86::KMOVWmk;
3107
3108 if (X86::FR32XRegClass.hasSubClassEq(RC))
3109 return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
3110 if (X86::FR64XRegClass.hasSubClassEq(RC))
3111 return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
3112 if (X86::VR128XRegClass.hasSubClassEq(RC) ||
3113 X86::VR256XRegClass.hasSubClassEq(RC) ||
3114 X86::VR512RegClass.hasSubClassEq(RC))
3115 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3116 }
3117
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3118 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3119 switch (RC->getSize()) {
Rafael Espindola6635f982010-07-12 03:43:04 +0000[diff] [blame]3120 default:
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3121 llvm_unreachable("Unknown spill size");
3122 case 1:
3123 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3124 if (TM.getSubtarget<X86Subtarget>().is64Bit())
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3125 // Copying to or from a physical H register on x86-64 requires a NOREX
3126 // move. Otherwise use a normal move.
3127 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3128 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3129 return load ? X86::MOV8rm : X86::MOV8mr;
3130 case 2:
3131 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3132 return load ? X86::MOV16rm : X86::MOV16mr;
3133 case 4:
3134 if (X86::GR32RegClass.hasSubClassEq(RC))
3135 return load ? X86::MOV32rm : X86::MOV32mr;
3136 if (X86::FR32RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3137 return load ?
3138 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3139 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3140 if (X86::RFP32RegClass.hasSubClassEq(RC))
3141 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3142 llvm_unreachable("Unknown 4-byte regclass");
3143 case 8:
3144 if (X86::GR64RegClass.hasSubClassEq(RC))
3145 return load ? X86::MOV64rm : X86::MOV64mr;
3146 if (X86::FR64RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3147 return load ?
3148 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3149 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3150 if (X86::VR64RegClass.hasSubClassEq(RC))
3151 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3152 if (X86::RFP64RegClass.hasSubClassEq(RC))
3153 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3154 llvm_unreachable("Unknown 8-byte regclass");
3155 case 10:
3156 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3157 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]3158 case 16: {
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3159 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3160 // If stack is realigned we can use aligned stores.
3161 if (isStackAligned)
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]3162 return load ?
3163 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
3164 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3165 else
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]3166 return load ?
3167 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
3168 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
3169 }
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]3170 case 32:
3171 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
3172 // If stack is realigned we can use aligned stores.
3173 if (isStackAligned)
3174 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
3175 else
3176 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3177 case 64:
3178 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3179 if (isStackAligned)
3180 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3181 else
3182 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3183 }
3184}
3185
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3186static unsigned getStoreRegOpcode(unsigned SrcReg,
3187 const TargetRegisterClass *RC,
3188 bool isStackAligned,
3189 TargetMachine &TM) {
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3190 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
3191}
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3192
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3193
3194static unsigned getLoadRegOpcode(unsigned DestReg,
3195 const TargetRegisterClass *RC,
3196 bool isStackAligned,
3197 const TargetMachine &TM) {
3198 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3199}
3200
3201void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3202 MachineBasicBlock::iterator MI,
3203 unsigned SrcReg, bool isKill, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]3204 const TargetRegisterClass *RC,
3205 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3206 const MachineFunction &MF = *MBB.getParent();
Jakob Stoklund Olesenc3c05ed2010-07-27 04:16:58 +0000[diff] [blame]3207 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
3208 "Stack slot too small for store");
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3209 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3210 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]3211 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3212 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]3213 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3214 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]3215 .addReg(SrcReg, getKillRegState(isKill));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3216}
3217
3218void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
3219 bool isKill,
3220 SmallVectorImpl<MachineOperand> &Addr,
3221 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3222 MachineInstr::mmo_iterator MMOBegin,
3223 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3224 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3225 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3226 bool isAligned = MMOBegin != MMOEnd &&
3227 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3228 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]3229 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]3230 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3231 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3232 MIB.addOperand(Addr[i]);
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]3233 MIB.addReg(SrcReg, getKillRegState(isKill));
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3234 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3235 NewMIs.push_back(MIB);
3236}
3237
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3238
3239void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3240 MachineBasicBlock::iterator MI,
3241 unsigned DestReg, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]3242 const TargetRegisterClass *RC,
3243 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3244 const MachineFunction &MF = *MBB.getParent();
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3245 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3246 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]3247 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3248 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]3249 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3250 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3251}
3252
3253void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]3254 SmallVectorImpl<MachineOperand> &Addr,
3255 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3256 MachineInstr::mmo_iterator MMOBegin,
3257 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3258 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3259 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3260 bool isAligned = MMOBegin != MMOEnd &&
3261 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3262 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]3263 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]3264 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3265 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3266 MIB.addOperand(Addr[i]);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3267 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3268 NewMIs.push_back(MIB);
3269}
3270
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3271bool X86InstrInfo::
3272analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3273 int &CmpMask, int &CmpValue) const {
3274 switch (MI->getOpcode()) {
3275 default: break;
3276 case X86::CMP64ri32:
3277 case X86::CMP64ri8:
3278 case X86::CMP32ri:
3279 case X86::CMP32ri8:
3280 case X86::CMP16ri:
3281 case X86::CMP16ri8:
3282 case X86::CMP8ri:
3283 SrcReg = MI->getOperand(0).getReg();
3284 SrcReg2 = 0;
3285 CmpMask = ~0;
3286 CmpValue = MI->getOperand(1).getImm();
3287 return true;
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3288 // A SUB can be used to perform comparison.
3289 case X86::SUB64rm:
3290 case X86::SUB32rm:
3291 case X86::SUB16rm:
3292 case X86::SUB8rm:
3293 SrcReg = MI->getOperand(1).getReg();
3294 SrcReg2 = 0;
3295 CmpMask = ~0;
3296 CmpValue = 0;
3297 return true;
3298 case X86::SUB64rr:
3299 case X86::SUB32rr:
3300 case X86::SUB16rr:
3301 case X86::SUB8rr:
3302 SrcReg = MI->getOperand(1).getReg();
3303 SrcReg2 = MI->getOperand(2).getReg();
3304 CmpMask = ~0;
3305 CmpValue = 0;
3306 return true;
3307 case X86::SUB64ri32:
3308 case X86::SUB64ri8:
3309 case X86::SUB32ri:
3310 case X86::SUB32ri8:
3311 case X86::SUB16ri:
3312 case X86::SUB16ri8:
3313 case X86::SUB8ri:
3314 SrcReg = MI->getOperand(1).getReg();
3315 SrcReg2 = 0;
3316 CmpMask = ~0;
3317 CmpValue = MI->getOperand(2).getImm();
3318 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3319 case X86::CMP64rr:
3320 case X86::CMP32rr:
3321 case X86::CMP16rr:
3322 case X86::CMP8rr:
3323 SrcReg = MI->getOperand(0).getReg();
3324 SrcReg2 = MI->getOperand(1).getReg();
3325 CmpMask = ~0;
3326 CmpValue = 0;
3327 return true;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3328 case X86::TEST8rr:
3329 case X86::TEST16rr:
3330 case X86::TEST32rr:
3331 case X86::TEST64rr:
3332 SrcReg = MI->getOperand(0).getReg();
3333 if (MI->getOperand(1).getReg() != SrcReg) return false;
3334 // Compare against zero.
3335 SrcReg2 = 0;
3336 CmpMask = ~0;
3337 CmpValue = 0;
3338 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3339 }
3340 return false;
3341}
3342
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3343/// isRedundantFlagInstr - check whether the first instruction, whose only
3344/// purpose is to update flags, can be made redundant.
3345/// CMPrr can be made redundant by SUBrr if the operands are the same.
3346/// This function can be extended later on.
3347/// SrcReg, SrcRegs: register operands for FlagI.
3348/// ImmValue: immediate for FlagI if it takes an immediate.
3349inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3350 unsigned SrcReg2, int ImmValue,
3351 MachineInstr *OI) {
3352 if (((FlagI->getOpcode() == X86::CMP64rr &&
3353 OI->getOpcode() == X86::SUB64rr) ||
3354 (FlagI->getOpcode() == X86::CMP32rr &&
3355 OI->getOpcode() == X86::SUB32rr)||
3356 (FlagI->getOpcode() == X86::CMP16rr &&
3357 OI->getOpcode() == X86::SUB16rr)||
3358 (FlagI->getOpcode() == X86::CMP8rr &&
3359 OI->getOpcode() == X86::SUB8rr)) &&
3360 ((OI->getOperand(1).getReg() == SrcReg &&
3361 OI->getOperand(2).getReg() == SrcReg2) ||
3362 (OI->getOperand(1).getReg() == SrcReg2 &&
3363 OI->getOperand(2).getReg() == SrcReg)))
3364 return true;
3365
3366 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3367 OI->getOpcode() == X86::SUB64ri32) ||
3368 (FlagI->getOpcode() == X86::CMP64ri8 &&
3369 OI->getOpcode() == X86::SUB64ri8) ||
3370 (FlagI->getOpcode() == X86::CMP32ri &&
3371 OI->getOpcode() == X86::SUB32ri) ||
3372 (FlagI->getOpcode() == X86::CMP32ri8 &&
3373 OI->getOpcode() == X86::SUB32ri8) ||
3374 (FlagI->getOpcode() == X86::CMP16ri &&
3375 OI->getOpcode() == X86::SUB16ri) ||
3376 (FlagI->getOpcode() == X86::CMP16ri8 &&
3377 OI->getOpcode() == X86::SUB16ri8) ||
3378 (FlagI->getOpcode() == X86::CMP8ri &&
3379 OI->getOpcode() == X86::SUB8ri)) &&
3380 OI->getOperand(1).getReg() == SrcReg &&
3381 OI->getOperand(2).getImm() == ImmValue)
3382 return true;
3383 return false;
3384}
3385
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3386/// isDefConvertible - check whether the definition can be converted
3387/// to remove a comparison against zero.
3388inline static bool isDefConvertible(MachineInstr *MI) {
3389 switch (MI->getOpcode()) {
3390 default: return false;
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]3391
3392 // The shift instructions only modify ZF if their shift count is non-zero.
3393 // N.B.: The processor truncates the shift count depending on the encoding.
3394 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3395 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3396 return getTruncatedShiftCount(MI, 2) != 0;
3397
3398 // Some left shift instructions can be turned into LEA instructions but only
3399 // if their flags aren't used. Avoid transforming such instructions.
3400 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3401 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3402 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3403 return ShAmt != 0;
3404 }
3405
3406 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3407 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3408 return getTruncatedShiftCount(MI, 3) != 0;
3409
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3410 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3411 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3412 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3413 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3414 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
Craig Topper5b08cf72012-12-17 04:55:07 +0000[diff] [blame]3415 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3416 case X86::DEC64_32r: case X86::DEC64_16r:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3417 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3418 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3419 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3420 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3421 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
Craig Topper5b08cf72012-12-17 04:55:07 +0000[diff] [blame]3422 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3423 case X86::INC64_32r: case X86::INC64_16r:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3424 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3425 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3426 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3427 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3428 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3429 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3430 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3431 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3432 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3433 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3434 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3435 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3436 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3437 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3438 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3439 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3440 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3441 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3442 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3443 case X86::ADC32ri: case X86::ADC32ri8:
3444 case X86::ADC32rr: case X86::ADC64ri32:
3445 case X86::ADC64ri8: case X86::ADC64rr:
3446 case X86::SBB32ri: case X86::SBB32ri8:
3447 case X86::SBB32rr: case X86::SBB64ri32:
3448 case X86::SBB64ri8: case X86::SBB64rr:
Craig Topperf3ff6ae2012-12-17 05:12:30 +0000[diff] [blame]3449 case X86::ANDN32rr: case X86::ANDN32rm:
3450 case X86::ANDN64rr: case X86::ANDN64rm:
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3451 case X86::BEXTR32rr: case X86::BEXTR64rr:
3452 case X86::BEXTR32rm: case X86::BEXTR64rm:
3453 case X86::BLSI32rr: case X86::BLSI32rm:
3454 case X86::BLSI64rr: case X86::BLSI64rm:
3455 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3456 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3457 case X86::BLSR32rr: case X86::BLSR32rm:
3458 case X86::BLSR64rr: case X86::BLSR64rm:
3459 case X86::BZHI32rr: case X86::BZHI32rm:
3460 case X86::BZHI64rr: case X86::BZHI64rm:
3461 case X86::LZCNT16rr: case X86::LZCNT16rm:
3462 case X86::LZCNT32rr: case X86::LZCNT32rm:
3463 case X86::LZCNT64rr: case X86::LZCNT64rm:
3464 case X86::POPCNT16rr:case X86::POPCNT16rm:
3465 case X86::POPCNT32rr:case X86::POPCNT32rm:
3466 case X86::POPCNT64rr:case X86::POPCNT64rm:
3467 case X86::TZCNT16rr: case X86::TZCNT16rm:
3468 case X86::TZCNT32rr: case X86::TZCNT32rm:
3469 case X86::TZCNT64rr: case X86::TZCNT64rm:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3470 return true;
3471 }
3472}
3473
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3474/// optimizeCompareInstr - Check if there exists an earlier instruction that
3475/// operates on the same source operands and sets flags in the same way as
3476/// Compare; remove Compare if possible.
3477bool X86InstrInfo::
3478optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3479 int CmpMask, int CmpValue,
3480 const MachineRegisterInfo *MRI) const {
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3481 // Check whether we can replace SUB with CMP.
3482 unsigned NewOpcode = 0;
3483 switch (CmpInstr->getOpcode()) {
3484 default: break;
3485 case X86::SUB64ri32:
3486 case X86::SUB64ri8:
3487 case X86::SUB32ri:
3488 case X86::SUB32ri8:
3489 case X86::SUB16ri:
3490 case X86::SUB16ri8:
3491 case X86::SUB8ri:
3492 case X86::SUB64rm:
3493 case X86::SUB32rm:
3494 case X86::SUB16rm:
3495 case X86::SUB8rm:
3496 case X86::SUB64rr:
3497 case X86::SUB32rr:
3498 case X86::SUB16rr:
3499 case X86::SUB8rr: {
3500 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3501 return false;
3502 // There is no use of the destination register, we can replace SUB with CMP.
3503 switch (CmpInstr->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]3504 default: llvm_unreachable("Unreachable!");
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3505 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3506 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3507 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3508 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3509 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3510 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3511 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3512 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3513 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3514 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3515 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3516 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3517 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3518 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3519 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3520 }
3521 CmpInstr->setDesc(get(NewOpcode));
3522 CmpInstr->RemoveOperand(0);
3523 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3524 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3525 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3526 return false;
3527 }
3528 }
3529
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3530 // Get the unique definition of SrcReg.
3531 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3532 if (!MI) return false;
3533
3534 // CmpInstr is the first instruction of the BB.
3535 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3536
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3537 // If we are comparing against zero, check whether we can use MI to update
3538 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3539 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3540 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3541 !isDefConvertible(MI)))
3542 return false;
3543
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3544 // We are searching for an earlier instruction that can make CmpInstr
3545 // redundant and that instruction will be saved in Sub.
3546 MachineInstr *Sub = NULL;
3547 const TargetRegisterInfo *TRI = &getRegisterInfo();
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3548
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3549 // We iterate backward, starting from the instruction before CmpInstr and
3550 // stop when reaching the definition of a source register or done with the BB.
3551 // RI points to the instruction before CmpInstr.
3552 // If the definition is in this basic block, RE points to the definition;
3553 // otherwise, RE is the rend of the basic block.
3554 MachineBasicBlock::reverse_iterator
3555 RI = MachineBasicBlock::reverse_iterator(I),
3556 RE = CmpInstr->getParent() == MI->getParent() ?
3557 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3558 CmpInstr->getParent()->rend();
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3559 MachineInstr *Movr0Inst = 0;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3560 for (; RI != RE; ++RI) {
3561 MachineInstr *Instr = &*RI;
3562 // Check whether CmpInstr can be made redundant by the current instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3563 if (!IsCmpZero &&
3564 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3565 Sub = Instr;
3566 break;
3567 }
3568
3569 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3570 Instr->readsRegister(X86::EFLAGS, TRI)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3571 // This instruction modifies or uses EFLAGS.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3572
3573 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3574 // They are safe to move up, if the definition to EFLAGS is dead and
3575 // earlier instructions do not read or write EFLAGS.
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]3576 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3577 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3578 Movr0Inst = Instr;
3579 continue;
3580 }
3581
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3582 // We can't remove CmpInstr.
3583 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3584 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3585 }
3586
3587 // Return false if no candidates exist.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3588 if (!IsCmpZero && !Sub)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3589 return false;
3590
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3591 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3592 Sub->getOperand(2).getReg() == SrcReg);
3593
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3594 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3595 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3596 // If we are done with the basic block, we need to check whether EFLAGS is
3597 // live-out.
3598 bool IsSafe = false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3599 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3600 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3601 for (++I; I != E; ++I) {
3602 const MachineInstr &Instr = *I;
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3603 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3604 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3605 // We should check the usage if this instruction uses and updates EFLAGS.
3606 if (!UseEFLAGS && ModifyEFLAGS) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3607 // It is safe to remove CmpInstr if EFLAGS is updated again.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3608 IsSafe = true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3609 break;
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3610 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3611 if (!UseEFLAGS && !ModifyEFLAGS)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3612 continue;
3613
3614 // EFLAGS is used by this instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3615 X86::CondCode OldCC;
3616 bool OpcIsSET = false;
3617 if (IsCmpZero || IsSwapped) {
3618 // We decode the condition code from opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3619 if (Instr.isBranch())
3620 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3621 else {
3622 OldCC = getCondFromSETOpc(Instr.getOpcode());
3623 if (OldCC != X86::COND_INVALID)
3624 OpcIsSET = true;
3625 else
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]3626 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3627 }
3628 if (OldCC == X86::COND_INVALID) return false;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3629 }
3630 if (IsCmpZero) {
3631 switch (OldCC) {
3632 default: break;
3633 case X86::COND_A: case X86::COND_AE:
3634 case X86::COND_B: case X86::COND_BE:
3635 case X86::COND_G: case X86::COND_GE:
3636 case X86::COND_L: case X86::COND_LE:
3637 case X86::COND_O: case X86::COND_NO:
3638 // CF and OF are used, we can't perform this optimization.
3639 return false;
3640 }
3641 } else if (IsSwapped) {
3642 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3643 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3644 // We swap the condition code and synthesize the new opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3645 X86::CondCode NewCC = getSwappedCondition(OldCC);
3646 if (NewCC == X86::COND_INVALID) return false;
3647
3648 // Synthesize the new opcode.
3649 bool HasMemoryOperand = Instr.hasOneMemOperand();
3650 unsigned NewOpc;
3651 if (Instr.isBranch())
3652 NewOpc = GetCondBranchFromCond(NewCC);
3653 else if(OpcIsSET)
3654 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3655 else {
3656 unsigned DstReg = Instr.getOperand(0).getReg();
3657 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3658 HasMemoryOperand);
3659 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3660
3661 // Push the MachineInstr to OpsToUpdate.
3662 // If it is safe to remove CmpInstr, the condition code of these
3663 // instructions will be modified.
3664 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3665 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3666 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3667 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3668 IsSafe = true;
3669 break;
3670 }
3671 }
3672
3673 // If EFLAGS is not killed nor re-defined, we should check whether it is
3674 // live-out. If it is live-out, do not optimize.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3675 if ((IsCmpZero || IsSwapped) && !IsSafe) {
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3676 MachineBasicBlock *MBB = CmpInstr->getParent();
3677 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3678 SE = MBB->succ_end(); SI != SE; ++SI)
3679 if ((*SI)->isLiveIn(X86::EFLAGS))
3680 return false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3681 }
3682
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3683 // The instruction to be updated is either Sub or MI.
3684 Sub = IsCmpZero ? MI : Sub;
David Majnemer5ba473a2013-05-18 01:02:03 +0000[diff] [blame]3685 // Move Movr0Inst to the appropriate place before Sub.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3686 if (Movr0Inst) {
David Majnemer5ba473a2013-05-18 01:02:03 +0000[diff] [blame]3687 // Look backwards until we find a def that doesn't use the current EFLAGS.
3688 Def = Sub;
3689 MachineBasicBlock::reverse_iterator
3690 InsertI = MachineBasicBlock::reverse_iterator(++Def),
3691 InsertE = Sub->getParent()->rend();
3692 for (; InsertI != InsertE; ++InsertI) {
3693 MachineInstr *Instr = &*InsertI;
3694 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3695 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3696 Sub->getParent()->remove(Movr0Inst);
3697 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3698 Movr0Inst);
3699 break;
3700 }
3701 }
3702 if (InsertI == InsertE)
3703 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3704 }
3705
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3706 // Make sure Sub instruction defines EFLAGS and mark the def live.
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3707 unsigned i = 0, e = Sub->getNumOperands();
3708 for (; i != e; ++i) {
3709 MachineOperand &MO = Sub->getOperand(i);
3710 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3711 MO.setIsDead(false);
3712 break;
3713 }
3714 }
3715 assert(i != e && "Unable to locate a def EFLAGS operand");
3716
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3717 CmpInstr->eraseFromParent();
3718
3719 // Modify the condition code of instructions in OpsToUpdate.
3720 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3721 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3722 return true;
3723}
3724
Manman Ren5759d012012-08-02 00:56:42 +0000[diff] [blame]3725/// optimizeLoadInstr - Try to remove the load by folding it to a register
3726/// operand at the use. We fold the load instructions if load defines a virtual
3727/// register, the virtual register is used once in the same BB, and the
3728/// instructions in-between do not load or store, and have no side effects.
3729MachineInstr* X86InstrInfo::
3730optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3731 unsigned &FoldAsLoadDefReg,
3732 MachineInstr *&DefMI) const {
3733 if (FoldAsLoadDefReg == 0)
3734 return 0;
3735 // To be conservative, if there exists another load, clear the load candidate.
3736 if (MI->mayLoad()) {
3737 FoldAsLoadDefReg = 0;
3738 return 0;
3739 }
3740
3741 // Check whether we can move DefMI here.
3742 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3743 assert(DefMI);
3744 bool SawStore = false;
3745 if (!DefMI->isSafeToMove(this, 0, SawStore))
3746 return 0;
3747
3748 // We try to commute MI if possible.
3749 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3750 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3751 // Collect information about virtual register operands of MI.
3752 unsigned SrcOperandId = 0;
3753 bool FoundSrcOperand = false;
3754 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3755 MachineOperand &MO = MI->getOperand(i);
3756 if (!MO.isReg())
3757 continue;
3758 unsigned Reg = MO.getReg();
3759 if (Reg != FoldAsLoadDefReg)
3760 continue;
3761 // Do not fold if we have a subreg use or a def or multiple uses.
3762 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3763 return 0;
3764
3765 SrcOperandId = i;
3766 FoundSrcOperand = true;
3767 }
3768 if (!FoundSrcOperand) return 0;
3769
3770 // Check whether we can fold the def into SrcOperandId.
3771 SmallVector<unsigned, 8> Ops;
3772 Ops.push_back(SrcOperandId);
3773 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3774 if (FoldMI) {
3775 FoldAsLoadDefReg = 0;
3776 return FoldMI;
3777 }
3778
3779 if (Idx == 1) {
3780 // MI was changed but it didn't help, commute it back!
3781 commuteInstruction(MI, false);
3782 return 0;
3783 }
3784
3785 // Check whether we can commute MI and enable folding.
3786 if (MI->isCommutable()) {
3787 MachineInstr *NewMI = commuteInstruction(MI, false);
3788 // Unable to commute.
3789 if (!NewMI) return 0;
3790 if (NewMI != MI) {
3791 // New instruction. It doesn't need to be kept.
3792 NewMI->eraseFromParent();
3793 return 0;
3794 }
3795 }
3796 }
3797 return 0;
3798}
3799
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3800/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3801/// instruction with two undef reads of the register being defined. This is
3802/// used for mapping:
3803/// %xmm4 = V_SET0
3804/// to:
3805/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3806///
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3807static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3808 const MCInstrDesc &Desc) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3809 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3810 unsigned Reg = MIB->getOperand(0).getReg();
3811 MIB->setDesc(Desc);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3812
3813 // MachineInstr::addOperand() will insert explicit operands before any
3814 // implicit operands.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3815 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3816 // But we don't trust that.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3817 assert(MIB->getOperand(1).getReg() == Reg &&
3818 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3819 return true;
3820}
3821
3822bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3823 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3824 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3825 switch (MI->getOpcode()) {
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3826 case X86::SETB_C8r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3827 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3828 case X86::SETB_C16r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3829 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3830 case X86::SETB_C32r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3831 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3832 case X86::SETB_C64r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3833 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3834 case X86::V_SET0:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3835 case X86::FsFLD0SS:
3836 case X86::FsFLD0SD:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3837 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]3838 case X86::AVX_SET0:
3839 assert(HasAVX && "AVX not supported");
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3840 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
Elena Demikhovskyf8f478b2013-08-25 12:54:30 +0000[diff] [blame]3841 case X86::AVX512_512_SET0:
3842 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3843 case X86::V_SETALLONES:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3844 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3845 case X86::AVX2_SETALLONES:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3846 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
Jakob Stoklund Olesen729abd32011-10-08 18:28:28 +0000[diff] [blame]3847 case X86::TEST8ri_NOREX:
3848 MI->setDesc(get(X86::TEST8ri));
3849 return true;
Elena Demikhovskyf8f478b2013-08-25 12:54:30 +0000[diff] [blame]3850 case X86::KSET0W: return Expand2AddrUndef(MIB, get(X86::KXORWrr));
3851 case X86::KSET1B:
3852 case X86::KSET1W: return Expand2AddrUndef(MIB, get(X86::KXNORWrr));
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3853 }
3854 return false;
3855}
3856
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3857static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3858 const SmallVectorImpl<MachineOperand> &MOs,
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3859 MachineInstr *MI,
3860 const TargetInstrInfo &TII) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3861 // Create the base instruction with the memory operand as the first part.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3862 // Omit the implicit operands, something BuildMI can't do.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3863 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3864 MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3865 MachineInstrBuilder MIB(MF, NewMI);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3866 unsigned NumAddrOps = MOs.size();
3867 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3868 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3869 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3870 addOffset(MIB, 0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3871
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3872 // Loop over the rest of the ri operands, converting them over.
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3873 unsigned NumOps = MI->getDesc().getNumOperands()-2;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3874 for (unsigned i = 0; i != NumOps; ++i) {
3875 MachineOperand &MO = MI->getOperand(i+2);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3876 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3877 }
3878 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3879 MachineOperand &MO = MI->getOperand(i);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3880 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3881 }
3882 return MIB;
3883}
3884
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3885static MachineInstr *FuseInst(MachineFunction &MF,
3886 unsigned Opcode, unsigned OpNo,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3887 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3888 MachineInstr *MI, const TargetInstrInfo &TII) {
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3889 // Omit the implicit operands, something BuildMI can't do.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3890 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3891 MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3892 MachineInstrBuilder MIB(MF, NewMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3893
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3894 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3895 MachineOperand &MO = MI->getOperand(i);
3896 if (i == OpNo) {
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3897 assert(MO.isReg() && "Expected to fold into reg operand!");
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3898 unsigned NumAddrOps = MOs.size();
3899 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3900 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3901 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3902 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3903 } else {
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3904 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3905 }
3906 }
3907 return MIB;
3908}
3909
3910static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3911 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3912 MachineInstr *MI) {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3913 MachineFunction &MF = *MI->getParent()->getParent();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3914 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3915
3916 unsigned NumAddrOps = MOs.size();
3917 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3918 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3919 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3920 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3921 return MIB.addImm(0);
3922}
3923
3924MachineInstr*
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3925X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3926 MachineInstr *MI, unsigned i,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3927 const SmallVectorImpl<MachineOperand> &MOs,
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3928 unsigned Size, unsigned Align) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3929 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
Preston Gurdd6be4bf2013-03-27 23:16:18 +0000[diff] [blame]3930 bool isCallRegIndirect = TM.getSubtarget<X86Subtarget>().callRegIndirect();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3931 bool isTwoAddrFold = false;
Preston Gurdd6be4bf2013-03-27 23:16:18 +0000[diff] [blame]3932
3933 // Atom favors register form of call. So, we do not fold loads into calls
3934 // when X86Subtarget is Atom.
3935 if (isCallRegIndirect &&
3936 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
3937 return NULL;
3938 }
3939
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3940 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3941 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]3942 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3943
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]3944 // FIXME: AsmPrinter doesn't know how to handle
3945 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3946 if (MI->getOpcode() == X86::ADD32ri &&
3947 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3948 return NULL;
3949
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3950 MachineInstr *NewMI = NULL;
3951 // Folding a memory location into the two-address part of a two-address
3952 // instruction is different than folding it other places. It requires
3953 // replacing the *two* registers with the memory location.
3954 if (isTwoAddr && NumOps >= 2 && i < 2 &&
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3955 MI->getOperand(0).isReg() &&
3956 MI->getOperand(1).isReg() &&
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3957 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3958 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3959 isTwoAddrFold = true;
3960 } else if (i == 0) { // If operand 0
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]3961 if (MI->getOpcode() == X86::MOV32r0) {
3962 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
3963 if (NewMI)
3964 return NewMI;
Craig Topperf9115972012-08-23 04:57:36 +0000[diff] [blame]3965 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3966
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3967 OpcodeTablePtr = &RegOp2MemOpTable0;
3968 } else if (i == 1) {
3969 OpcodeTablePtr = &RegOp2MemOpTable1;
3970 } else if (i == 2) {
3971 OpcodeTablePtr = &RegOp2MemOpTable2;
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]3972 } else if (i == 3) {
3973 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3974 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3975
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3976 // If table selected...
3977 if (OpcodeTablePtr) {
3978 // Find the Opcode to fuse
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3979 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3980 OpcodeTablePtr->find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3981 if (I != OpcodeTablePtr->end()) {
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3982 unsigned Opcode = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]3983 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3984 if (Align < MinAlign)
3985 return NULL;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3986 bool NarrowToMOV32rm = false;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3987 if (Size) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]3988 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3989 if (Size < RCSize) {
3990 // Check if it's safe to fold the load. If the size of the object is
3991 // narrower than the load width, then it's not.
3992 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3993 return NULL;
3994 // If this is a 64-bit load, but the spill slot is 32, then we can do
3995 // a 32-bit load which is implicitly zero-extended. This likely is due
3996 // to liveintervalanalysis remat'ing a load from stack slot.
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3997 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3998 return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3999 Opcode = X86::MOV32rm;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]4000 NarrowToMOV32rm = true;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4001 }
4002 }
4003
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4004 if (isTwoAddrFold)
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4005 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4006 else
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4007 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]4008
4009 if (NarrowToMOV32rm) {
4010 // If this is the special case where we use a MOV32rm to load a 32-bit
4011 // value and zero-extend the top bits. Change the destination register
4012 // to a 32-bit one.
4013 unsigned DstReg = NewMI->getOperand(0).getReg();
4014 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4015 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]4016 X86::sub_32bit));
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]4017 else
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]4018 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]4019 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4020 return NewMI;
4021 }
4022 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4023
4024 // No fusion
Jakob Stoklund Olesen51702ec2010-07-09 20:43:09 +0000[diff] [blame]4025 if (PrintFailedFusing && !MI->isCopy())
David Greened589daf2010-01-05 01:29:29 +0000[diff] [blame]4026 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4027 return NULL;
4028}
4029
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4030/// hasPartialRegUpdate - Return true for all instructions that only update
4031/// the first 32 or 64-bits of the destination register and leave the rest
4032/// unmodified. This can be used to avoid folding loads if the instructions
4033/// only update part of the destination register, and the non-updated part is
4034/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4035/// instructions breaks the partial register dependency and it can improve
4036/// performance. e.g.:
4037///
4038/// movss (%rdi), %xmm0
4039/// cvtss2sd %xmm0, %xmm0
4040///
4041/// Instead of
4042/// cvtss2sd (%rdi), %xmm0
4043///
Bruno Cardoso Lopes7b435682011-09-15 23:04:24 +0000[diff] [blame]4044/// FIXME: This should be turned into a TSFlags.
4045///
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4046static bool hasPartialRegUpdate(unsigned Opcode) {
4047 switch (Opcode) {
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]4048 case X86::CVTSI2SSrr:
4049 case X86::CVTSI2SS64rr:
4050 case X86::CVTSI2SDrr:
4051 case X86::CVTSI2SD64rr:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4052 case X86::CVTSD2SSrr:
4053 case X86::Int_CVTSD2SSrr:
4054 case X86::CVTSS2SDrr:
4055 case X86::Int_CVTSS2SDrr:
4056 case X86::RCPSSr:
4057 case X86::RCPSSr_Int:
4058 case X86::ROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]4059 case X86::ROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4060 case X86::ROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]4061 case X86::ROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4062 case X86::RSQRTSSr:
4063 case X86::RSQRTSSr_Int:
4064 case X86::SQRTSSr:
4065 case X86::SQRTSSr_Int:
4066 // AVX encoded versions
4067 case X86::VCVTSD2SSrr:
4068 case X86::Int_VCVTSD2SSrr:
4069 case X86::VCVTSS2SDrr:
4070 case X86::Int_VCVTSS2SDrr:
Elena Demikhovsky534015e2013-09-02 07:12:29 +0000[diff] [blame]4071 case X86::VCVTSD2SSZrr:
4072 case X86::VCVTSS2SDZrr:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4073 case X86::VRCPSSr:
4074 case X86::VROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]4075 case X86::VROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4076 case X86::VROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]4077 case X86::VROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4078 case X86::VRSQRTSSr:
4079 case X86::VSQRTSSr:
4080 return true;
4081 }
4082
4083 return false;
4084}
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4085
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]4086/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
4087/// instructions we would like before a partial register update.
4088unsigned X86InstrInfo::
4089getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
4090 const TargetRegisterInfo *TRI) const {
4091 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
4092 return 0;
4093
4094 // If MI is marked as reading Reg, the partial register update is wanted.
4095 const MachineOperand &MO = MI->getOperand(0);
4096 unsigned Reg = MO.getReg();
4097 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4098 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
4099 return 0;
4100 } else {
4101 if (MI->readsRegister(Reg, TRI))
4102 return 0;
4103 }
4104
4105 // If any of the preceding 16 instructions are reading Reg, insert a
4106 // dependency breaking instruction. The magic number is based on a few
4107 // Nehalem experiments.
4108 return 16;
4109}
4110
4111void X86InstrInfo::
4112breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
4113 const TargetRegisterInfo *TRI) const {
4114 unsigned Reg = MI->getOperand(OpNum).getReg();
4115 if (X86::VR128RegClass.contains(Reg)) {
4116 // These instructions are all floating point domain, so xorps is the best
4117 // choice.
4118 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
4119 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
4120 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
4121 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4122 } else if (X86::VR256RegClass.contains(Reg)) {
4123 // Use vxorps to clear the full ymm register.
4124 // It wants to read and write the xmm sub-register.
4125 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4126 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
4127 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
4128 .addReg(Reg, RegState::ImplicitDefine);
4129 } else
4130 return;
4131 MI->addRegisterKilled(Reg, TRI, true);
4132}
4133
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]4134MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4135 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]4136 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]4137 int FrameIndex) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4138 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4139 if (NoFusing) return NULL;
4140
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4141 // Unless optimizing for size, don't fold to avoid partial
4142 // register update stalls
Bill Wendling698e84f2012-12-30 10:32:01 +0000[diff] [blame]4143 if (!MF.getFunction()->getAttributes().
4144 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4145 hasPartialRegUpdate(MI->getOpcode()))
4146 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]4147
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]4148 const MachineFrameInfo *MFI = MF.getFrameInfo();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4149 unsigned Size = MFI->getObjectSize(FrameIndex);
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]4150 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4151 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4152 unsigned NewOpc = 0;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4153 unsigned RCSize = 0;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4154 switch (MI->getOpcode()) {
4155 default: return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4156 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
Dan Gohman887dd1c2010-05-18 21:42:03 +0000[diff] [blame]4157 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
4158 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
4159 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4160 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4161 // Check if it's safe to fold the load. If the size of the object is
4162 // narrower than the load width, then it's not.
4163 if (Size < RCSize)
4164 return NULL;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4165 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4166 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4167 MI->getOperand(1).ChangeToImmediate(0);
4168 } else if (Ops.size() != 1)
4169 return NULL;
4170
4171 SmallVector<MachineOperand,4> MOs;
4172 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4173 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4174}
4175
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]4176MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4177 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]4178 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]4179 MachineInstr *LoadMI) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4180 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4181 if (NoFusing) return NULL;
4182
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4183 // Unless optimizing for size, don't fold to avoid partial
4184 // register update stalls
Bill Wendling698e84f2012-12-30 10:32:01 +0000[diff] [blame]4185 if (!MF.getFunction()->getAttributes().
4186 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4187 hasPartialRegUpdate(MI->getOpcode()))
4188 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]4189
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]4190 // Determine the alignment of the load.
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]4191 unsigned Alignment = 0;
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]4192 if (LoadMI->hasOneMemOperand())
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]4193 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4194 else
4195 switch (LoadMI->getOpcode()) {
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]4196 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4197 case X86::AVX_SET0:
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4198 Alignment = 32;
4199 break;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4200 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4201 case X86::V_SETALLONES:
4202 Alignment = 16;
4203 break;
4204 case X86::FsFLD0SD:
4205 Alignment = 8;
4206 break;
4207 case X86::FsFLD0SS:
4208 Alignment = 4;
4209 break;
4210 default:
Eli Friedman87ef3872011-06-10 01:13:01 +0000[diff] [blame]4211 return 0;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4212 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4213 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4214 unsigned NewOpc = 0;
4215 switch (MI->getOpcode()) {
4216 default: return NULL;
4217 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4218 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
4219 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
4220 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4221 }
4222 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4223 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4224 MI->getOperand(1).ChangeToImmediate(0);
4225 } else if (Ops.size() != 1)
4226 return NULL;
4227
Jakob Stoklund Olesen9c473e42010-08-11 23:08:22 +0000[diff] [blame]4228 // Make sure the subregisters match.
4229 // Otherwise we risk changing the size of the load.
4230 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4231 return NULL;
4232
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4233 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4234 switch (LoadMI->getOpcode()) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4235 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4236 case X86::V_SETALLONES:
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]4237 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4238 case X86::AVX_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4239 case X86::FsFLD0SD:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4240 case X86::FsFLD0SS: {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4241 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4242 // Create a constant-pool entry and operands to load from it.
4243
Dan Gohman772952f2010-03-09 03:01:40 +0000[diff] [blame]4244 // Medium and large mode can't fold loads this way.
4245 if (TM.getCodeModel() != CodeModel::Small &&
4246 TM.getCodeModel() != CodeModel::Kernel)
4247 return NULL;
4248
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4249 // x86-32 PIC requires a PIC base register for constant pools.
4250 unsigned PICBase = 0;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4251 if (TM.getRelocationModel() == Reloc::PIC_) {
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]4252 if (TM.getSubtarget<X86Subtarget>().is64Bit())
4253 PICBase = X86::RIP;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4254 else
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4255 // FIXME: PICBase = getGlobalBaseReg(&MF);
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]4256 // This doesn't work for several reasons.
4257 // 1. GlobalBaseReg may have been spilled.
4258 // 2. It may not be live at MI.
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4259 return NULL;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4260 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4261
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4262 // Create a constant-pool entry.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4263 MachineConstantPool &MCP = *MF.getConstantPool();
Chris Lattner229907c2011-07-18 04:54:35 +0000[diff] [blame]4264 Type *Ty;
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4265 unsigned Opc = LoadMI->getOpcode();
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4266 if (Opc == X86::FsFLD0SS)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4267 Ty = Type::getFloatTy(MF.getFunction()->getContext());
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4268 else if (Opc == X86::FsFLD0SD)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4269 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4270 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
Craig Toppera4c5a472012-01-13 06:12:41 +0000[diff] [blame]4271 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4272 else
4273 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]4274
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]4275 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]4276 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4277 Constant::getNullValue(Ty);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4278 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4279
4280 // Create operands to load from the constant pool entry.
4281 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4282 MOs.push_back(MachineOperand::CreateImm(1));
4283 MOs.push_back(MachineOperand::CreateReg(0, false));
4284 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]4285 MOs.push_back(MachineOperand::CreateReg(0, false));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4286 break;
4287 }
4288 default: {
Manman Ren5b462822012-11-27 18:09:26 +0000[diff] [blame]4289 if ((LoadMI->getOpcode() == X86::MOVSSrm ||
4290 LoadMI->getOpcode() == X86::VMOVSSrm) &&
4291 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4292 > 4)
4293 // These instructions only load 32 bits, we can't fold them if the
4294 // destination register is wider than 32 bits (4 bytes).
4295 return NULL;
4296 if ((LoadMI->getOpcode() == X86::MOVSDrm ||
4297 LoadMI->getOpcode() == X86::VMOVSDrm) &&
4298 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4299 > 8)
4300 // These instructions only load 64 bits, we can't fold them if the
4301 // destination register is wider than 64 bits (8 bytes).
4302 return NULL;
4303
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4304 // Folding a normal load. Just copy the load's address operands.
4305 unsigned NumOps = LoadMI->getDesc().getNumOperands();
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4306 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4307 MOs.push_back(LoadMI->getOperand(i));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4308 break;
4309 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4310 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4311 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4312}
4313
4314
Dan Gohman33332bc2008-10-16 01:49:15 +0000[diff] [blame]4315bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4316 const SmallVectorImpl<unsigned> &Ops) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4317 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4318 if (NoFusing) return 0;
4319
4320 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4321 switch (MI->getOpcode()) {
4322 default: return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4323 case X86::TEST8rr:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4324 case X86::TEST16rr:
4325 case X86::TEST32rr:
4326 case X86::TEST64rr:
4327 return true;
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]4328 case X86::ADD32ri:
4329 // FIXME: AsmPrinter doesn't know how to handle
4330 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4331 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4332 return false;
4333 break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4334 }
4335 }
4336
4337 if (Ops.size() != 1)
4338 return false;
4339
4340 unsigned OpNum = Ops[0];
4341 unsigned Opc = MI->getOpcode();
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]4342 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4343 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4344 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4345
4346 // Folding a memory location into the two-address part of a two-address
4347 // instruction is different than folding it other places. It requires
4348 // replacing the *two* registers with the memory location.
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4349 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4350 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4351 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4352 } else if (OpNum == 0) { // If operand 0
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]4353 if (Opc == X86::MOV32r0)
4354 return true;
4355
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4356 OpcodeTablePtr = &RegOp2MemOpTable0;
4357 } else if (OpNum == 1) {
4358 OpcodeTablePtr = &RegOp2MemOpTable1;
4359 } else if (OpNum == 2) {
4360 OpcodeTablePtr = &RegOp2MemOpTable2;
Craig Topper7573c8f2012-08-31 22:12:16 +0000[diff] [blame]4361 } else if (OpNum == 3) {
4362 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4363 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4364
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]4365 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4366 return true;
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]4367 return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4368}
4369
4370bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4371 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4372 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4373 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4374 MemOp2RegOpTable.find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4375 if (I == MemOp2RegOpTable.end())
4376 return false;
4377 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4378 unsigned Index = I->second.second & TB_INDEX_MASK;
4379 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4380 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4381 if (UnfoldLoad && !FoldedLoad)
4382 return false;
4383 UnfoldLoad &= FoldedLoad;
4384 if (UnfoldStore && !FoldedStore)
4385 return false;
4386 UnfoldStore &= FoldedStore;
4387
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4388 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4389 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4390 if (!MI->hasOneMemOperand() &&
4391 RC == &X86::VR128RegClass &&
4392 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4393 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4394 // conservatively assume the address is unaligned. That's bad for
4395 // performance.
4396 return false;
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4397 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4398 SmallVector<MachineOperand,2> BeforeOps;
4399 SmallVector<MachineOperand,2> AfterOps;
4400 SmallVector<MachineOperand,4> ImpOps;
4401 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4402 MachineOperand &Op = MI->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4403 if (i >= Index && i < Index + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4404 AddrOps.push_back(Op);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4405 else if (Op.isReg() && Op.isImplicit())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4406 ImpOps.push_back(Op);
4407 else if (i < Index)
4408 BeforeOps.push_back(Op);
4409 else if (i > Index)
4410 AfterOps.push_back(Op);
4411 }
4412
4413 // Emit the load instruction.
4414 if (UnfoldLoad) {
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4415 std::pair<MachineInstr::mmo_iterator,
4416 MachineInstr::mmo_iterator> MMOs =
4417 MF.extractLoadMemRefs(MI->memoperands_begin(),
4418 MI->memoperands_end());
4419 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4420 if (UnfoldStore) {
4421 // Address operands cannot be marked isKill.
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4422 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4423 MachineOperand &MO = NewMIs[0]->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4424 if (MO.isReg())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4425 MO.setIsKill(false);
4426 }
4427 }
4428 }
4429
4430 // Emit the data processing instruction.
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4431 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]4432 MachineInstrBuilder MIB(MF, DataMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4433
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4434 if (FoldedStore)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4435 MIB.addReg(Reg, RegState::Define);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4436 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4437 MIB.addOperand(BeforeOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4438 if (FoldedLoad)
4439 MIB.addReg(Reg);
4440 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4441 MIB.addOperand(AfterOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4442 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4443 MachineOperand &MO = ImpOps[i];
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4444 MIB.addReg(MO.getReg(),
4445 getDefRegState(MO.isDef()) |
4446 RegState::Implicit |
4447 getKillRegState(MO.isKill()) |
Evan Cheng0dc101b2009-06-30 08:49:04 +0000[diff] [blame]4448 getDeadRegState(MO.isDead()) |
4449 getUndefRegState(MO.isUndef()));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4450 }
4451 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4452 switch (DataMI->getOpcode()) {
4453 default: break;
4454 case X86::CMP64ri32:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4455 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4456 case X86::CMP32ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4457 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4458 case X86::CMP16ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4459 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4460 case X86::CMP8ri: {
4461 MachineOperand &MO0 = DataMI->getOperand(0);
4462 MachineOperand &MO1 = DataMI->getOperand(1);
4463 if (MO1.getImm() == 0) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4464 unsigned NewOpc;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4465 switch (DataMI->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4466 default: llvm_unreachable("Unreachable!");
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4467 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4468 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4469 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4470 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4471 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4472 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4473 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4474 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4475 DataMI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4476 MO1.ChangeToRegister(MO0.getReg(), false);
4477 }
4478 }
4479 }
4480 NewMIs.push_back(DataMI);
4481
4482 // Emit the store instruction.
4483 if (UnfoldStore) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4484 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4485 std::pair<MachineInstr::mmo_iterator,
4486 MachineInstr::mmo_iterator> MMOs =
4487 MF.extractStoreMemRefs(MI->memoperands_begin(),
4488 MI->memoperands_end());
4489 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4490 }
4491
4492 return true;
4493}
4494
4495bool
4496X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4497 SmallVectorImpl<SDNode*> &NewNodes) const {
Dan Gohman17059682008-07-17 19:10:17 +0000[diff] [blame]4498 if (!N->isMachineOpcode())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4499 return false;
4500
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4501 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4502 MemOp2RegOpTable.find(N->getMachineOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4503 if (I == MemOp2RegOpTable.end())
4504 return false;
4505 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4506 unsigned Index = I->second.second & TB_INDEX_MASK;
4507 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4508 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4509 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4510 MachineFunction &MF = DAG.getMachineFunction();
4511 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4512 unsigned NumDefs = MCID.NumDefs;
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4513 std::vector<SDValue> AddrOps;
4514 std::vector<SDValue> BeforeOps;
4515 std::vector<SDValue> AfterOps;
Andrew Trickef9de2a2013-05-25 02:42:55 +0000[diff] [blame]4516 SDLoc dl(N);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4517 unsigned NumOps = N->getNumOperands();
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]4518 for (unsigned i = 0; i != NumOps-1; ++i) {
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4519 SDValue Op = N->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4520 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4521 AddrOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4522 else if (i < Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4523 BeforeOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4524 else if (i > Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4525 AfterOps.push_back(Op);
4526 }
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4527 SDValue Chain = N->getOperand(NumOps-1);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4528 AddrOps.push_back(Chain);
4529
4530 // Emit the load instruction.
4531 SDNode *Load = 0;
4532 if (FoldedLoad) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4533 EVT VT = *RC->vt_begin();
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4534 std::pair<MachineInstr::mmo_iterator,
4535 MachineInstr::mmo_iterator> MMOs =
4536 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4537 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4538 if (!(*MMOs.first) &&
4539 RC == &X86::VR128RegClass &&
4540 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4541 // Do not introduce a slow unaligned load.
4542 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4543 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4544 bool isAligned = (*MMOs.first) &&
4545 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4546 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4547 VT, MVT::Other, AddrOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4548 NewNodes.push_back(Load);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4549
4550 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4551 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4552 }
4553
4554 // Emit the data processing instruction.
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4555 std::vector<EVT> VTs;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4556 const TargetRegisterClass *DstRC = 0;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4557 if (MCID.getNumDefs() > 0) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4558 DstRC = getRegClass(MCID, 0, &RI, MF);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4559 VTs.push_back(*DstRC->vt_begin());
4560 }
4561 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4562 EVT VT = N->getValueType(i);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4563 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4564 VTs.push_back(VT);
4565 }
4566 if (Load)
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4567 BeforeOps.push_back(SDValue(Load, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4568 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4569 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4570 NewNodes.push_back(NewNode);
4571
4572 // Emit the store instruction.
4573 if (FoldedStore) {
4574 AddrOps.pop_back();
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4575 AddrOps.push_back(SDValue(NewNode, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4576 AddrOps.push_back(Chain);
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4577 std::pair<MachineInstr::mmo_iterator,
4578 MachineInstr::mmo_iterator> MMOs =
4579 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4580 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4581 if (!(*MMOs.first) &&
4582 RC == &X86::VR128RegClass &&
4583 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4584 // Do not introduce a slow unaligned store.
4585 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4586 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4587 bool isAligned = (*MMOs.first) &&
4588 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4589 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4590 isAligned, TM),
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4591 dl, MVT::Other, AddrOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4592 NewNodes.push_back(Store);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4593
4594 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4595 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4596 }
4597
4598 return true;
4599}
4600
4601unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4602 bool UnfoldLoad, bool UnfoldStore,
4603 unsigned *LoadRegIndex) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4604 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4605 MemOp2RegOpTable.find(Opc);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4606 if (I == MemOp2RegOpTable.end())
4607 return 0;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4608 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4609 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4610 if (UnfoldLoad && !FoldedLoad)
4611 return 0;
4612 if (UnfoldStore && !FoldedStore)
4613 return 0;
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4614 if (LoadRegIndex)
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4615 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4616 return I->second.first;
4617}
4618
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4619bool
4620X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4621 int64_t &Offset1, int64_t &Offset2) const {
4622 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4623 return false;
4624 unsigned Opc1 = Load1->getMachineOpcode();
4625 unsigned Opc2 = Load2->getMachineOpcode();
4626 switch (Opc1) {
4627 default: return false;
4628 case X86::MOV8rm:
4629 case X86::MOV16rm:
4630 case X86::MOV32rm:
4631 case X86::MOV64rm:
4632 case X86::LD_Fp32m:
4633 case X86::LD_Fp64m:
4634 case X86::LD_Fp80m:
4635 case X86::MOVSSrm:
4636 case X86::MOVSDrm:
4637 case X86::MMX_MOVD64rm:
4638 case X86::MMX_MOVQ64rm:
4639 case X86::FsMOVAPSrm:
4640 case X86::FsMOVAPDrm:
4641 case X86::MOVAPSrm:
4642 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4643 case X86::MOVAPDrm:
4644 case X86::MOVDQArm:
4645 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4646 // AVX load instructions
4647 case X86::VMOVSSrm:
4648 case X86::VMOVSDrm:
4649 case X86::FsVMOVAPSrm:
4650 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4651 case X86::VMOVAPSrm:
4652 case X86::VMOVUPSrm:
4653 case X86::VMOVAPDrm:
4654 case X86::VMOVDQArm:
4655 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4656 case X86::VMOVAPSYrm:
4657 case X86::VMOVUPSYrm:
4658 case X86::VMOVAPDYrm:
4659 case X86::VMOVDQAYrm:
4660 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4661 break;
4662 }
4663 switch (Opc2) {
4664 default: return false;
4665 case X86::MOV8rm:
4666 case X86::MOV16rm:
4667 case X86::MOV32rm:
4668 case X86::MOV64rm:
4669 case X86::LD_Fp32m:
4670 case X86::LD_Fp64m:
4671 case X86::LD_Fp80m:
4672 case X86::MOVSSrm:
4673 case X86::MOVSDrm:
4674 case X86::MMX_MOVD64rm:
4675 case X86::MMX_MOVQ64rm:
4676 case X86::FsMOVAPSrm:
4677 case X86::FsMOVAPDrm:
4678 case X86::MOVAPSrm:
4679 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4680 case X86::MOVAPDrm:
4681 case X86::MOVDQArm:
4682 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4683 // AVX load instructions
4684 case X86::VMOVSSrm:
4685 case X86::VMOVSDrm:
4686 case X86::FsVMOVAPSrm:
4687 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4688 case X86::VMOVAPSrm:
4689 case X86::VMOVUPSrm:
4690 case X86::VMOVAPDrm:
4691 case X86::VMOVDQArm:
4692 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4693 case X86::VMOVAPSYrm:
4694 case X86::VMOVUPSYrm:
4695 case X86::VMOVAPDYrm:
4696 case X86::VMOVDQAYrm:
4697 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4698 break;
4699 }
4700
4701 // Check if chain operands and base addresses match.
4702 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4703 Load1->getOperand(5) != Load2->getOperand(5))
4704 return false;
4705 // Segment operands should match as well.
4706 if (Load1->getOperand(4) != Load2->getOperand(4))
4707 return false;
4708 // Scale should be 1, Index should be Reg0.
4709 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4710 Load1->getOperand(2) == Load2->getOperand(2)) {
4711 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4712 return false;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4713
4714 // Now let's examine the displacements.
4715 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4716 isa<ConstantSDNode>(Load2->getOperand(3))) {
4717 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4718 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4719 return true;
4720 }
4721 }
4722 return false;
4723}
4724
4725bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4726 int64_t Offset1, int64_t Offset2,
4727 unsigned NumLoads) const {
4728 assert(Offset2 > Offset1);
4729 if ((Offset2 - Offset1) / 8 > 64)
4730 return false;
4731
4732 unsigned Opc1 = Load1->getMachineOpcode();
4733 unsigned Opc2 = Load2->getMachineOpcode();
4734 if (Opc1 != Opc2)
4735 return false; // FIXME: overly conservative?
4736
4737 switch (Opc1) {
4738 default: break;
4739 case X86::LD_Fp32m:
4740 case X86::LD_Fp64m:
4741 case X86::LD_Fp80m:
4742 case X86::MMX_MOVD64rm:
4743 case X86::MMX_MOVQ64rm:
4744 return false;
4745 }
4746
4747 EVT VT = Load1->getValueType(0);
4748 switch (VT.getSimpleVT().SimpleTy) {
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4749 default:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4750 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4751 // have 16 of them to play with.
4752 if (TM.getSubtargetImpl()->is64Bit()) {
4753 if (NumLoads >= 3)
4754 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4755 } else if (NumLoads) {
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4756 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4757 }
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4758 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4759 case MVT::i8:
4760 case MVT::i16:
4761 case MVT::i32:
4762 case MVT::i64:
Evan Cheng16cf9342010-01-22 23:49:11 +0000[diff] [blame]4763 case MVT::f32:
4764 case MVT::f64:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4765 if (NumLoads)
4766 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4767 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4768 }
4769
4770 return true;
4771}
4772
Andrew Trick47740de2013-06-23 09:00:28 +0000[diff] [blame]4773bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First,
4774 MachineInstr *Second) const {
4775 // Check if this processor supports macro-fusion. Since this is a minor
4776 // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
4777 // proxy for SandyBridge+.
4778 if (!TM.getSubtarget<X86Subtarget>().hasAVX())
4779 return false;
4780
4781 enum {
4782 FuseTest,
4783 FuseCmp,
4784 FuseInc
4785 } FuseKind;
4786
4787 switch(Second->getOpcode()) {
4788 default:
4789 return false;
4790 case X86::JE_4:
4791 case X86::JNE_4:
4792 case X86::JL_4:
4793 case X86::JLE_4:
4794 case X86::JG_4:
4795 case X86::JGE_4:
4796 FuseKind = FuseInc;
4797 break;
4798 case X86::JB_4:
4799 case X86::JBE_4:
4800 case X86::JA_4:
4801 case X86::JAE_4:
4802 FuseKind = FuseCmp;
4803 break;
4804 case X86::JS_4:
4805 case X86::JNS_4:
4806 case X86::JP_4:
4807 case X86::JNP_4:
4808 case X86::JO_4:
4809 case X86::JNO_4:
4810 FuseKind = FuseTest;
4811 break;
4812 }
4813 switch (First->getOpcode()) {
4814 default:
4815 return false;
4816 case X86::TEST8rr:
4817 case X86::TEST16rr:
4818 case X86::TEST32rr:
4819 case X86::TEST64rr:
4820 case X86::TEST8ri:
4821 case X86::TEST16ri:
4822 case X86::TEST32ri:
4823 case X86::TEST32i32:
4824 case X86::TEST64i32:
4825 case X86::TEST64ri32:
4826 case X86::TEST8rm:
4827 case X86::TEST16rm:
4828 case X86::TEST32rm:
4829 case X86::TEST64rm:
4830 case X86::AND16i16:
4831 case X86::AND16ri:
4832 case X86::AND16ri8:
4833 case X86::AND16rm:
4834 case X86::AND16rr:
4835 case X86::AND32i32:
4836 case X86::AND32ri:
4837 case X86::AND32ri8:
4838 case X86::AND32rm:
4839 case X86::AND32rr:
4840 case X86::AND64i32:
4841 case X86::AND64ri32:
4842 case X86::AND64ri8:
4843 case X86::AND64rm:
4844 case X86::AND64rr:
4845 case X86::AND8i8:
4846 case X86::AND8ri:
4847 case X86::AND8rm:
4848 case X86::AND8rr:
4849 return true;
4850 case X86::CMP16i16:
4851 case X86::CMP16ri:
4852 case X86::CMP16ri8:
4853 case X86::CMP16rm:
4854 case X86::CMP16rr:
4855 case X86::CMP32i32:
4856 case X86::CMP32ri:
4857 case X86::CMP32ri8:
4858 case X86::CMP32rm:
4859 case X86::CMP32rr:
4860 case X86::CMP64i32:
4861 case X86::CMP64ri32:
4862 case X86::CMP64ri8:
4863 case X86::CMP64rm:
4864 case X86::CMP64rr:
4865 case X86::CMP8i8:
4866 case X86::CMP8ri:
4867 case X86::CMP8rm:
4868 case X86::CMP8rr:
4869 case X86::ADD16i16:
4870 case X86::ADD16ri:
4871 case X86::ADD16ri8:
4872 case X86::ADD16ri8_DB:
4873 case X86::ADD16ri_DB:
4874 case X86::ADD16rm:
4875 case X86::ADD16rr:
4876 case X86::ADD16rr_DB:
4877 case X86::ADD32i32:
4878 case X86::ADD32ri:
4879 case X86::ADD32ri8:
4880 case X86::ADD32ri8_DB:
4881 case X86::ADD32ri_DB:
4882 case X86::ADD32rm:
4883 case X86::ADD32rr:
4884 case X86::ADD32rr_DB:
4885 case X86::ADD64i32:
4886 case X86::ADD64ri32:
4887 case X86::ADD64ri32_DB:
4888 case X86::ADD64ri8:
4889 case X86::ADD64ri8_DB:
4890 case X86::ADD64rm:
4891 case X86::ADD64rr:
4892 case X86::ADD64rr_DB:
4893 case X86::ADD8i8:
4894 case X86::ADD8mi:
4895 case X86::ADD8mr:
4896 case X86::ADD8ri:
4897 case X86::ADD8rm:
4898 case X86::ADD8rr:
4899 case X86::SUB16i16:
4900 case X86::SUB16ri:
4901 case X86::SUB16ri8:
4902 case X86::SUB16rm:
4903 case X86::SUB16rr:
4904 case X86::SUB32i32:
4905 case X86::SUB32ri:
4906 case X86::SUB32ri8:
4907 case X86::SUB32rm:
4908 case X86::SUB32rr:
4909 case X86::SUB64i32:
4910 case X86::SUB64ri32:
4911 case X86::SUB64ri8:
4912 case X86::SUB64rm:
4913 case X86::SUB64rr:
4914 case X86::SUB8i8:
4915 case X86::SUB8ri:
4916 case X86::SUB8rm:
4917 case X86::SUB8rr:
4918 return FuseKind == FuseCmp || FuseKind == FuseInc;
4919 case X86::INC16r:
4920 case X86::INC32r:
4921 case X86::INC64_16r:
4922 case X86::INC64_32r:
4923 case X86::INC64r:
4924 case X86::INC8r:
4925 case X86::DEC16r:
4926 case X86::DEC32r:
4927 case X86::DEC64_16r:
4928 case X86::DEC64_32r:
4929 case X86::DEC64r:
4930 case X86::DEC8r:
4931 return FuseKind == FuseInc;
4932 }
4933}
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4934
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4935bool X86InstrInfo::
Owen Anderson4f6bf042008-08-14 22:49:33 +0000[diff] [blame]4936ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4937 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4938 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]4939 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4940 return true;
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4941 Cond[0].setImm(GetOppositeBranchCondition(CC));
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4942 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4943}
4944
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4945bool X86InstrInfo::
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4946isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4947 // FIXME: Return false for x87 stack register classes for now. We can't
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4948 // allow any loads of these registers before FpGet_ST0_80.
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4949 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4950 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4951}
4952
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4953/// getGlobalBaseReg - Return a virtual register initialized with the
4954/// the global base register value. Output instructions required to
4955/// initialize the register in the function entry block, if necessary.
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4956///
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4957/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4958///
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4959unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4960 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4961 "X86-64 PIC uses RIP relative addressing");
4962
4963 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4964 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4965 if (GlobalBaseReg != 0)
4966 return GlobalBaseReg;
4967
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4968 // Create the register. The code to initialize it is inserted
4969 // later, by the CGBR pass (below).
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4970 MachineRegisterInfo &RegInfo = MF->getRegInfo();
Jakob Stoklund Olesen38dcd592012-05-20 18:43:00 +0000[diff] [blame]4971 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4972 X86FI->setGlobalBaseReg(GlobalBaseReg);
4973 return GlobalBaseReg;
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4974}
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4975
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4976// These are the replaceable SSE instructions. Some of these have Int variants
4977// that we don't include here. We don't want to replace instructions selected
4978// by intrinsics.
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4979static const uint16_t ReplaceableInstrs[][3] = {
Bruno Cardoso Lopes1401e042010-08-12 02:08:52 +0000[diff] [blame]4980 //PackedSingle PackedDouble PackedInt
Jakob Stoklund Olesendbff4e82010-03-30 22:46:53 +0000[diff] [blame]4981 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4982 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4983 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4984 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4985 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4986 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4987 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4988 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4989 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4990 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4991 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4992 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4993 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4994 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4995 // AVX 128-bit support
4996 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4997 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4998 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4999 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
5000 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
5001 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
5002 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
5003 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
5004 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
5005 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
5006 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
5007 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]5008 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
5009 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]5010 // AVX 256-bit support
5011 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
5012 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
5013 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
5014 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
5015 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]5016 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
5017};
5018
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]5019static const uint16_t ReplaceableInstrsAVX2[][3] = {
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]5020 //PackedSingle PackedDouble PackedInt
Craig Topperf87a2be2011-11-09 09:37:21 +0000[diff] [blame]5021 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
5022 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
5023 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
5024 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
5025 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
5026 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
5027 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
Craig Topper12b72de2011-11-29 05:37:58 +0000[diff] [blame]5028 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
5029 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
5030 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
5031 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
5032 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
5033 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
5034 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5035};
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]5036
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5037// FIXME: Some shuffle and unpack instructions have equivalents in different
5038// domains, but they require a bit more work than just switching opcodes.
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]5039
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]5040static const uint16_t *lookup(unsigned opcode, unsigned domain) {
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]5041 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5042 if (ReplaceableInstrs[i][domain-1] == opcode)
5043 return ReplaceableInstrs[i];
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]5044 return 0;
5045}
5046
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]5047static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]5048 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
5049 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
5050 return ReplaceableInstrsAVX2[i];
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5051 return 0;
5052}
5053
5054std::pair<uint16_t, uint16_t>
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]5055X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5056 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]5057 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]5058 uint16_t validDomains = 0;
5059 if (domain && lookup(MI->getOpcode(), domain))
5060 validDomains = 0xe;
5061 else if (domain && lookupAVX2(MI->getOpcode(), domain))
5062 validDomains = hasAVX2 ? 0xe : 0x6;
5063 return std::make_pair(domain, validDomains);
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5064}
5065
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]5066void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5067 assert(Domain>0 && Domain<4 && "Invalid execution domain");
5068 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5069 assert(dom && "Not an SSE instruction");
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]5070 const uint16_t *table = lookup(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]5071 if (!table) { // try the other table
5072 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
5073 "256-bit vector operations only available in AVX2");
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]5074 table = lookupAVX2(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]5075 }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5076 assert(table && "Cannot change domain");
5077 MI->setDesc(get(table[Domain-1]));
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]5078}
Chris Lattner6a5e7062010-04-26 23:37:21 +0000[diff] [blame]5079
5080/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
5081void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
5082 NopInst.setOpcode(X86::NOOP);
5083}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5084
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]5085bool X86InstrInfo::isHighLatencyDef(int opc) const {
5086 switch (opc) {
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5087 default: return false;
5088 case X86::DIVSDrm:
5089 case X86::DIVSDrm_Int:
5090 case X86::DIVSDrr:
5091 case X86::DIVSDrr_Int:
5092 case X86::DIVSSrm:
5093 case X86::DIVSSrm_Int:
5094 case X86::DIVSSrr:
5095 case X86::DIVSSrr_Int:
5096 case X86::SQRTPDm:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5097 case X86::SQRTPDr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5098 case X86::SQRTPSm:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5099 case X86::SQRTPSr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5100 case X86::SQRTSDm:
5101 case X86::SQRTSDm_Int:
5102 case X86::SQRTSDr:
5103 case X86::SQRTSDr_Int:
5104 case X86::SQRTSSm:
5105 case X86::SQRTSSm_Int:
5106 case X86::SQRTSSr:
5107 case X86::SQRTSSr_Int:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5108 // AVX instructions with high latency
5109 case X86::VDIVSDrm:
5110 case X86::VDIVSDrm_Int:
5111 case X86::VDIVSDrr:
5112 case X86::VDIVSDrr_Int:
5113 case X86::VDIVSSrm:
5114 case X86::VDIVSSrm_Int:
5115 case X86::VDIVSSrr:
5116 case X86::VDIVSSrr_Int:
5117 case X86::VSQRTPDm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5118 case X86::VSQRTPDr:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5119 case X86::VSQRTPSm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5120 case X86::VSQRTPSr:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5121 case X86::VSQRTSDm:
5122 case X86::VSQRTSDm_Int:
5123 case X86::VSQRTSDr:
5124 case X86::VSQRTSSm:
5125 case X86::VSQRTSSm_Int:
5126 case X86::VSQRTSSr:
Elena Demikhovsky402ee642013-09-02 07:41:01 +0000[diff] [blame]5127 case X86::VSQRTPDZrm:
5128 case X86::VSQRTPDZrr:
5129 case X86::VSQRTPSZrm:
5130 case X86::VSQRTPSZrr:
5131 case X86::VSQRTSDZm:
5132 case X86::VSQRTSDZm_Int:
5133 case X86::VSQRTSDZr:
5134 case X86::VSQRTSSZm_Int:
5135 case X86::VSQRTSSZr:
5136 case X86::VSQRTSSZm:
5137 case X86::VDIVSDZrm:
5138 case X86::VDIVSDZrr:
5139 case X86::VDIVSSZrm:
5140 case X86::VDIVSSZrr:
Elena Demikhovsky534015e2013-09-02 07:12:29 +0000[diff] [blame]5141
5142 case X86::VGATHERQPSZrm:
5143 case X86::VGATHERQPDZrm:
5144 case X86::VGATHERDPDZrm:
5145 case X86::VGATHERDPSZrm:
5146 case X86::VPGATHERQDZrm:
5147 case X86::VPGATHERQQZrm:
5148 case X86::VPGATHERDDZrm:
Elena Demikhovsky402ee642013-09-02 07:41:01 +0000[diff] [blame]5149 case X86::VPGATHERDQZrm:
5150 case X86::VSCATTERQPDZmr:
5151 case X86::VSCATTERQPSZmr:
5152 case X86::VSCATTERDPDZmr:
5153 case X86::VSCATTERDPSZmr:
5154 case X86::VPSCATTERQDZmr:
5155 case X86::VPSCATTERQQZmr:
5156 case X86::VPSCATTERDDZmr:
5157 case X86::VPSCATTERDQZmr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5158 return true;
5159 }
5160}
5161
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]5162bool X86InstrInfo::
5163hasHighOperandLatency(const InstrItineraryData *ItinData,
5164 const MachineRegisterInfo *MRI,
5165 const MachineInstr *DefMI, unsigned DefIdx,
5166 const MachineInstr *UseMI, unsigned UseIdx) const {
5167 return isHighLatencyDef(DefMI->getOpcode());
5168}
5169
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5170namespace {
5171 /// CGBR - Create Global Base Reg pass. This initializes the PIC
5172 /// global base register for x86-32.
5173 struct CGBR : public MachineFunctionPass {
5174 static char ID;
Owen Andersona7aed182010-08-06 18:33:48 +0000[diff] [blame]5175 CGBR() : MachineFunctionPass(ID) {}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5176
5177 virtual bool runOnMachineFunction(MachineFunction &MF) {
5178 const X86TargetMachine *TM =
5179 static_cast<const X86TargetMachine *>(&MF.getTarget());
5180
5181 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
5182 "X86-64 PIC uses RIP relative addressing");
5183
5184 // Only emit a global base reg in PIC mode.
5185 if (TM->getRelocationModel() != Reloc::PIC_)
5186 return false;
5187
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]5188 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
5189 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5190
5191 // If we didn't need a GlobalBaseReg, don't insert code.
5192 if (GlobalBaseReg == 0)
5193 return false;
5194
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5195 // Insert the set of GlobalBaseReg into the first MBB of the function
5196 MachineBasicBlock &FirstMBB = MF.front();
5197 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
5198 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
5199 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5200 const X86InstrInfo *TII = TM->getInstrInfo();
5201
5202 unsigned PC;
5203 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]5204 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5205 else
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]5206 PC = GlobalBaseReg;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]5207
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5208 // Operand of MovePCtoStack is completely ignored by asm printer. It's
5209 // only used in JIT code emission as displacement to pc.
5210 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]5211
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5212 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
5213 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
5214 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5215 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
5216 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
5217 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
5218 X86II::MO_GOT_ABSOLUTE_ADDRESS);
5219 }
5220
5221 return true;
5222 }
5223
5224 virtual const char *getPassName() const {
5225 return "X86 PIC Global Base Reg Initialization";
5226 }
5227
5228 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
5229 AU.setPreservesCFG();
5230 MachineFunctionPass::getAnalysisUsage(AU);
5231 }
5232 };
5233}
5234
5235char CGBR::ID = 0;
5236FunctionPass*
5237llvm::createGlobalBaseRegPass() { return new CGBR(); }
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]5238
5239namespace {
5240 struct LDTLSCleanup : public MachineFunctionPass {
5241 static char ID;
5242 LDTLSCleanup() : MachineFunctionPass(ID) {}
5243
5244 virtual bool runOnMachineFunction(MachineFunction &MF) {
5245 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
5246 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
5247 // No point folding accesses if there isn't at least two.
5248 return false;
5249 }
5250
5251 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
5252 return VisitNode(DT->getRootNode(), 0);
5253 }
5254
5255 // Visit the dominator subtree rooted at Node in pre-order.
5256 // If TLSBaseAddrReg is non-null, then use that to replace any
5257 // TLS_base_addr instructions. Otherwise, create the register
5258 // when the first such instruction is seen, and then use it
5259 // as we encounter more instructions.
5260 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
5261 MachineBasicBlock *BB = Node->getBlock();
5262 bool Changed = false;
5263
5264 // Traverse the current block.
5265 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
5266 ++I) {
5267 switch (I->getOpcode()) {
5268 case X86::TLS_base_addr32:
5269 case X86::TLS_base_addr64:
5270 if (TLSBaseAddrReg)
5271 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
5272 else
5273 I = SetRegister(I, &TLSBaseAddrReg);
5274 Changed = true;
5275 break;
5276 default:
5277 break;
5278 }
5279 }
5280
5281 // Visit the children of this block in the dominator tree.
5282 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
5283 I != E; ++I) {
5284 Changed |= VisitNode(*I, TLSBaseAddrReg);
5285 }
5286
5287 return Changed;
5288 }
5289
5290 // Replace the TLS_base_addr instruction I with a copy from
5291 // TLSBaseAddrReg, returning the new instruction.
5292 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
5293 unsigned TLSBaseAddrReg) {
5294 MachineFunction *MF = I->getParent()->getParent();
5295 const X86TargetMachine *TM =
5296 static_cast<const X86TargetMachine *>(&MF->getTarget());
5297 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5298 const X86InstrInfo *TII = TM->getInstrInfo();
5299
5300 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
5301 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
5302 TII->get(TargetOpcode::COPY),
5303 is64Bit ? X86::RAX : X86::EAX)
5304 .addReg(TLSBaseAddrReg);
5305
5306 // Erase the TLS_base_addr instruction.
5307 I->eraseFromParent();
5308
5309 return Copy;
5310 }
5311
5312 // Create a virtal register in *TLSBaseAddrReg, and populate it by
5313 // inserting a copy instruction after I. Returns the new instruction.
5314 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
5315 MachineFunction *MF = I->getParent()->getParent();
5316 const X86TargetMachine *TM =
5317 static_cast<const X86TargetMachine *>(&MF->getTarget());
5318 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5319 const X86InstrInfo *TII = TM->getInstrInfo();
5320
5321 // Create a virtual register for the TLS base address.
5322 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5323 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
5324 ? &X86::GR64RegClass
5325 : &X86::GR32RegClass);
5326
5327 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
5328 MachineInstr *Next = I->getNextNode();
5329 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
5330 TII->get(TargetOpcode::COPY),
5331 *TLSBaseAddrReg)
5332 .addReg(is64Bit ? X86::RAX : X86::EAX);
5333
5334 return Copy;
5335 }
5336
5337 virtual const char *getPassName() const {
5338 return "Local Dynamic TLS Access Clean-up";
5339 }
5340
5341 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
5342 AU.setPreservesCFG();
5343 AU.addRequired<MachineDominatorTree>();
5344 MachineFunctionPass::getAnalysisUsage(AU);
5345 }
5346 };
5347}
5348
5349char LDTLSCleanup::ID = 0;
5350FunctionPass*
5351llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }