Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / be133a8757103c430a36f15d6210f99a0206921f / . / llvm / lib / Target / X86 / X86InstrInfo.cpp
blob: c4c090b5c953c8f6a17d9130654275137a0c89ca [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 },
378 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]379 };
380
381 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]382 unsigned RegOp = OpTbl0[i].RegOp;
383 unsigned MemOp = OpTbl0[i].MemOp;
384 unsigned Flags = OpTbl0[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]385 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
386 RegOp, MemOp, TB_INDEX_0 | Flags);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]387 }
388
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]389 static const X86OpTblEntry OpTbl1[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]390 { X86::CMP16rr, X86::CMP16rm, 0 },
391 { X86::CMP32rr, X86::CMP32rm, 0 },
392 { X86::CMP64rr, X86::CMP64rm, 0 },
393 { X86::CMP8rr, X86::CMP8rm, 0 },
394 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
395 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
396 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
397 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
398 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
399 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
400 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
401 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
402 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
403 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
404 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
405 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]406 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
407 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
408 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
409 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
410 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
411 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
412 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
413 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]414 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
415 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]416 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
417 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]418 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
419 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
420 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
421 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
422 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
423 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
424 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
425 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]426 { X86::MOV16rr, X86::MOV16rm, 0 },
427 { X86::MOV32rr, X86::MOV32rm, 0 },
428 { X86::MOV64rr, X86::MOV64rm, 0 },
429 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
430 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
431 { X86::MOV8rr, X86::MOV8rm, 0 },
432 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
433 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]434 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
435 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
436 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
437 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]438 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
439 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
440 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
441 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
442 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
443 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
444 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
445 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
446 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
447 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]448 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
449 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
450 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
451 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
452 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
453 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
454 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]455 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
456 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
457 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]458 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
459 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
460 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
461 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
462 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
463 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
464 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
465 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
466 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
467 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]468 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]469 { X86::SQRTSDr, X86::SQRTSDm, 0 },
470 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
471 { X86::SQRTSSr, X86::SQRTSSm, 0 },
472 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
473 { X86::TEST16rr, X86::TEST16rm, 0 },
474 { X86::TEST32rr, X86::TEST32rm, 0 },
475 { X86::TEST64rr, X86::TEST64rm, 0 },
476 { X86::TEST8rr, X86::TEST8rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]477 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]478 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
479 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]480 // AVX 128-bit versions of foldable instructions
481 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
482 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]483 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
484 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]485 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
486 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
Pete Cooper8bbce762012-06-14 22:12:58 +0000[diff] [blame]487 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]488 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
489 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
490 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
491 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
492 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
493 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
494 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
495 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
496 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]497 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
498 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
499 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
500 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
501 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
502 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
503 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
504 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
505 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
506 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
507 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
508 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
Craig Topperb2922162012-12-26 02:14:19 +0000[diff] [blame]509 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]510 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
511 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
512 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
513 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]514 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
515 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
516 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
517 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
518 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
519 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
520 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
521 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
522 { X86::VRCPPSr, X86::VRCPPSm, 0 },
523 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, 0 },
524 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
525 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, 0 },
526 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]527 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]528 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]529 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]530 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
531
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]532 // AVX 256-bit foldable instructions
533 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
534 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]535 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]536 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]537 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]538 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
539 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]540
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]541 // AVX2 foldable instructions
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]542 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
543 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
544 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
545 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
546 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
547 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
548 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
549 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, 0 },
550 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]551 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]552 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]553 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
554 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
Michael Liao2de86af2012-09-26 08:24:51 +0000[diff] [blame]555
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]556 // BMI/BMI2/LZCNT/POPCNT foldable instructions
557 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
558 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
559 { X86::BLSI32rr, X86::BLSI32rm, 0 },
560 { X86::BLSI64rr, X86::BLSI64rm, 0 },
561 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
562 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
563 { X86::BLSR32rr, X86::BLSR32rm, 0 },
564 { X86::BLSR64rr, X86::BLSR64rm, 0 },
565 { X86::BZHI32rr, X86::BZHI32rm, 0 },
566 { X86::BZHI64rr, X86::BZHI64rm, 0 },
567 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
568 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
569 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
570 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
571 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
572 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
Michael Liao2de86af2012-09-26 08:24:51 +0000[diff] [blame]573 { X86::RORX32ri, X86::RORX32mi, 0 },
574 { X86::RORX64ri, X86::RORX64mi, 0 },
Michael Liao2b425e12012-09-26 08:26:25 +0000[diff] [blame]575 { X86::SARX32rr, X86::SARX32rm, 0 },
576 { X86::SARX64rr, X86::SARX64rm, 0 },
577 { X86::SHRX32rr, X86::SHRX32rm, 0 },
578 { X86::SHRX64rr, X86::SHRX64rm, 0 },
579 { X86::SHLX32rr, X86::SHLX32rm, 0 },
580 { X86::SHLX64rr, X86::SHLX64rm, 0 },
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]581 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
582 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
583 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]584 };
585
586 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]587 unsigned RegOp = OpTbl1[i].RegOp;
588 unsigned MemOp = OpTbl1[i].MemOp;
589 unsigned Flags = OpTbl1[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]590 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
591 RegOp, MemOp,
592 // Index 1, folded load
593 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]594 }
595
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]596 static const X86OpTblEntry OpTbl2[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]597 { X86::ADC32rr, X86::ADC32rm, 0 },
598 { X86::ADC64rr, X86::ADC64rm, 0 },
599 { X86::ADD16rr, X86::ADD16rm, 0 },
600 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
601 { X86::ADD32rr, X86::ADD32rm, 0 },
602 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
603 { X86::ADD64rr, X86::ADD64rm, 0 },
604 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
605 { X86::ADD8rr, X86::ADD8rm, 0 },
606 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
607 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
608 { X86::ADDSDrr, X86::ADDSDrm, 0 },
609 { X86::ADDSSrr, X86::ADDSSrm, 0 },
610 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
611 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
612 { X86::AND16rr, X86::AND16rm, 0 },
613 { X86::AND32rr, X86::AND32rm, 0 },
614 { X86::AND64rr, X86::AND64rm, 0 },
615 { X86::AND8rr, X86::AND8rm, 0 },
616 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
617 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
618 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
619 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]620 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
621 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
622 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
623 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]624 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
625 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
626 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
627 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
628 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
629 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
630 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
631 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
632 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
633 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
634 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
635 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
636 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
637 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
638 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
639 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
640 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
641 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
642 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
643 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
644 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
645 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
646 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
647 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
648 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
649 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
650 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
651 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
652 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
653 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
654 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
655 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
656 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
657 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
658 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
659 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
660 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
661 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
662 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
663 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
664 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
665 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
666 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
667 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
668 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
669 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
670 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
671 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
672 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
673 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
674 { X86::CMPSDrr, X86::CMPSDrm, 0 },
675 { X86::CMPSSrr, X86::CMPSSrm, 0 },
676 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
677 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
678 { X86::DIVSDrr, X86::DIVSDrm, 0 },
679 { X86::DIVSSrr, X86::DIVSSrm, 0 },
680 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
681 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
682 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
683 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
684 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
685 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
686 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
687 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
688 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
689 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
690 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
691 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
692 { X86::IMUL16rr, X86::IMUL16rm, 0 },
693 { X86::IMUL32rr, X86::IMUL32rm, 0 },
694 { X86::IMUL64rr, X86::IMUL64rm, 0 },
695 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
696 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
Manman Ren959acb12012-08-13 18:29:41 +0000[diff] [blame]697 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
698 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
699 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
700 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
701 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
702 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]703 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]704 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]705 { X86::MAXSDrr, X86::MAXSDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]706 { X86::MAXSSrr, X86::MAXSSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]707 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]708 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]709 { X86::MINSDrr, X86::MINSDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]710 { X86::MINSSrr, X86::MINSSrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]711 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]712 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
713 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
714 { X86::MULSDrr, X86::MULSDrm, 0 },
715 { X86::MULSSrr, X86::MULSSrm, 0 },
716 { X86::OR16rr, X86::OR16rm, 0 },
717 { X86::OR32rr, X86::OR32rm, 0 },
718 { X86::OR64rr, X86::OR64rm, 0 },
719 { X86::OR8rr, X86::OR8rm, 0 },
720 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
721 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
722 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
723 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]724 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]725 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
726 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
727 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
728 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
729 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
730 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]731 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
732 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]733 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]734 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]735 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
736 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
737 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
738 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]739 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]740 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
741 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]742 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]743 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
744 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
745 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]746 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]747 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]748 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
749 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]750 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]751 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]752 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]753 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]754 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]755 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]756 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
757 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
758 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
759 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
760 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
Benjamin Kramer4669d182012-12-21 14:04:55 +0000[diff] [blame]761 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
762 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
763 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
764 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
765 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
766 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
767 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
768 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]769 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]770 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]771 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
772 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
773 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
774 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
775 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
776 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
777 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]778 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
779 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
780 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
781 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]782 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
783 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
784 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
785 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
786 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
787 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
788 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
789 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
790 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
791 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
792 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
793 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
794 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
795 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
796 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
797 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
798 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
799 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
800 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
801 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
802 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
803 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
804 { X86::SBB32rr, X86::SBB32rm, 0 },
805 { X86::SBB64rr, X86::SBB64rm, 0 },
806 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
807 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
808 { X86::SUB16rr, X86::SUB16rm, 0 },
809 { X86::SUB32rr, X86::SUB32rm, 0 },
810 { X86::SUB64rr, X86::SUB64rm, 0 },
811 { X86::SUB8rr, X86::SUB8rm, 0 },
812 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
813 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
814 { X86::SUBSDrr, X86::SUBSDrm, 0 },
815 { X86::SUBSSrr, X86::SUBSSrm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]816 // FIXME: TEST*rr -> swapped operand of TEST*mr.
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]817 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
818 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
819 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
820 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
821 { X86::XOR16rr, X86::XOR16rm, 0 },
822 { X86::XOR32rr, X86::XOR32rm, 0 },
823 { X86::XOR64rr, X86::XOR64rm, 0 },
824 { X86::XOR8rr, X86::XOR8rm, 0 },
825 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]826 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
827 // AVX 128-bit versions of foldable instructions
828 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
829 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
830 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
831 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
832 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
833 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
834 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
835 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
836 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
837 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
Craig Toppercaef1c52012-12-26 00:35:47 +0000[diff] [blame]838 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
839 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]840 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
841 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]842 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
843 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
844 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]845 { X86::VADDPDrr, X86::VADDPDrm, 0 },
846 { X86::VADDPSrr, X86::VADDPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]847 { X86::VADDSDrr, X86::VADDSDrm, 0 },
848 { X86::VADDSSrr, X86::VADDSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]849 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
850 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
851 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
852 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
853 { X86::VANDPDrr, X86::VANDPDrm, 0 },
854 { X86::VANDPSrr, X86::VANDPSrm, 0 },
855 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
856 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
857 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
858 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
859 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
860 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]861 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
862 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]863 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
864 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]865 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
866 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
867 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
868 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
869 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
870 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
871 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
872 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
873 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
874 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]875 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
876 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
877 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
878 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]879 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
880 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]881 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]882 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]883 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]884 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]885 { X86::VMINPDrr, X86::VMINPDrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]886 { X86::VMINPSrr, X86::VMINPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]887 { X86::VMINSDrr, X86::VMINSDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]888 { X86::VMINSSrr, X86::VMINSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]889 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
890 { X86::VMULPDrr, X86::VMULPDrm, 0 },
891 { X86::VMULPSrr, X86::VMULPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]892 { X86::VMULSDrr, X86::VMULSDrm, 0 },
893 { X86::VMULSSrr, X86::VMULSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]894 { X86::VORPDrr, X86::VORPDrm, 0 },
895 { X86::VORPSrr, X86::VORPSrm, 0 },
896 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
897 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
898 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
899 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
900 { X86::VPADDBrr, X86::VPADDBrm, 0 },
901 { X86::VPADDDrr, X86::VPADDDrm, 0 },
902 { X86::VPADDQrr, X86::VPADDQrm, 0 },
903 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
904 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
905 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
906 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
907 { X86::VPADDWrr, X86::VPADDWrm, 0 },
908 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
909 { X86::VPANDNrr, X86::VPANDNrm, 0 },
910 { X86::VPANDrr, X86::VPANDrm, 0 },
911 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
912 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
913 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
914 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
915 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
916 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
917 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
918 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
919 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
920 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
921 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
922 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
923 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
924 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
925 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
926 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
927 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
928 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
929 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
930 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
931 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
932 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
933 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
934 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
935 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
936 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
937 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
938 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
939 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
940 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
941 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
942 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
943 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
944 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
945 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
946 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
947 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
948 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
949 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
950 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
951 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
952 { X86::VPORrr, X86::VPORrm, 0 },
953 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
954 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
955 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
956 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
957 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
958 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
959 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
960 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
961 { X86::VPSRADrr, X86::VPSRADrm, 0 },
962 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
963 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
964 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
965 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
966 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
967 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
968 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
969 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
970 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
971 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
972 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
973 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
974 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
975 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
976 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
977 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
978 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
979 { X86::VPXORrr, X86::VPXORrm, 0 },
980 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
981 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
982 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
983 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]984 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
985 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]986 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
987 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
988 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
989 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
990 { X86::VXORPDrr, X86::VXORPDrm, 0 },
991 { X86::VXORPSrr, X86::VXORPSrm, 0 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]992 // AVX 256-bit foldable instructions
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]993 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
994 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
995 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
996 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
997 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
998 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
999 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
1000 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1001 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1002 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1003 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1004 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1005 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1006 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1007 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1008 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1009 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1010 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1011 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1012 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1013 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1014 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1015 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1016 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1017 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1018 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1019 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1020 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1021 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1022 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1023 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1024 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1025 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1026 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1027 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1028 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1029 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1030 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1031 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1032 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1033 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1034 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1035 // AVX2 foldable instructions
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1036 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1037 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1038 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1039 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1040 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1041 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1042 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1043 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1044 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1045 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1046 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1047 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1048 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1049 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1050 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1051 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1052 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1053 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1054 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1055 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1056 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1057 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1058 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1059 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1060 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1061 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1062 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1063 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1064 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1065 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1066 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1067 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
1068 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1069 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
1070 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1071 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1072 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1073 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1074 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1075 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1076 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1077 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1078 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1079 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1080 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1081 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1082 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1083 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1084 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1085 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1086 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1087 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1088 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1089 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1090 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1091 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1092 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1093 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1094 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1095 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1096 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1097 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1098 { X86::VPORYrr, X86::VPORYrm, 0 },
1099 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1100 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1101 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1102 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1103 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1104 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1105 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1106 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1107 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1108 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1109 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1110 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1111 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1112 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1113 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1114 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1115 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1116 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1117 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1118 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1119 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1120 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1121 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1122 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1123 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1124 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1125 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1126 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1127 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1128 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1129 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1130 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1131 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1132 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1133 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1134 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1135 { X86::VPXORYrr, X86::VPXORYrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1136 // FIXME: add AVX 256-bit foldable instructions
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1137
1138 // FMA4 foldable patterns
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1139 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, 0 },
1140 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1141 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1142 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1143 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1144 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1145 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, 0 },
1146 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1147 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1148 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1149 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1150 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1151 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, 0 },
1152 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1153 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1154 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1155 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1156 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1157 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, 0 },
1158 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1159 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1160 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1161 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1162 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1163 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1164 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1165 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1166 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1167 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1168 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1169 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1170 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
Michael Liaof9f7b552012-09-26 08:22:37 +0000[diff] [blame]1171
1172 // BMI/BMI2 foldable instructions
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]1173 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1174 { X86::ANDN64rr, X86::ANDN64rm, 0 },
Michael Liaof9f7b552012-09-26 08:22:37 +0000[diff] [blame]1175 { X86::MULX32rr, X86::MULX32rm, 0 },
1176 { X86::MULX64rr, X86::MULX64rm, 0 },
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]1177 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1178 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1179 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1180 { X86::PEXT64rr, X86::PEXT64rm, 0 },
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]1181
1182 // AVX-512 foldable instructions
1183 { X86::VPERMPDZri, X86::VPERMPDZmi, 0 },
1184 { X86::VPERMPSZrr, X86::VPERMPSZrm, 0 },
1185 { X86::VPERMI2Drr, X86::VPERMI2Drm, 0 },
1186 { X86::VPERMI2Qrr, X86::VPERMI2Qrm, 0 },
1187 { X86::VPERMI2PSrr, X86::VPERMI2PSrm, 0 },
1188 { X86::VPERMI2PDrr, X86::VPERMI2PDrm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1189 };
1190
1191 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]1192 unsigned RegOp = OpTbl2[i].RegOp;
1193 unsigned MemOp = OpTbl2[i].MemOp;
1194 unsigned Flags = OpTbl2[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1195 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1196 RegOp, MemOp,
1197 // Index 2, folded load
1198 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1199 }
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1200
1201 static const X86OpTblEntry OpTbl3[] = {
1202 // FMA foldable instructions
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1203 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1204 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1205 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1206 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1207 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1208 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1209 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, 0 },
1210 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1211
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1212 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1213 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1214 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1215 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1216 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1217 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1218 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1219 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1220 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1221 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1222 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1223 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1224
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1225 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1226 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1227 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1228 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1229 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1230 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1231 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, 0 },
1232 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1233
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1234 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1235 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1236 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1237 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1238 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1239 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1240 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1241 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1242 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1243 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1244 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1245 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1246
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1247 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1248 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1249 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1250 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1251 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1252 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1253 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, 0 },
1254 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1255
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1256 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1257 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1258 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1259 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1260 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1261 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1262 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1263 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1264 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1265 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1266 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1267 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1268
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1269 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1270 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1271 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1272 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1273 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1274 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1275 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, 0 },
1276 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, 0 },
Craig Topper2e127b52012-06-01 05:48:39 +0000[diff] [blame]1277
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1278 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1279 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1280 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1281 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1282 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1283 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1284 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1285 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1286 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1287 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1288 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1289 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1290
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1291 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1292 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1293 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1294 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1295 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1296 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1297 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1298 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1299 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1300 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1301 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1302 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1303
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1304 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1305 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1306 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1307 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1308 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1309 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1310 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1311 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1312 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1313 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1314 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1315 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1316
1317 // FMA4 foldable patterns
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1318 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, 0 },
1319 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1320 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1321 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1322 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1323 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1324 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, 0 },
1325 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1326 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1327 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1328 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1329 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1330 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, 0 },
1331 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1332 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1333 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1334 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1335 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1336 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, 0 },
1337 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1338 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1339 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1340 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1341 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1342 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1343 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1344 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1345 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1346 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1347 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1348 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1349 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1350 };
1351
1352 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1353 unsigned RegOp = OpTbl3[i].RegOp;
1354 unsigned MemOp = OpTbl3[i].MemOp;
1355 unsigned Flags = OpTbl3[i].Flags;
1356 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1357 RegOp, MemOp,
1358 // Index 3, folded load
1359 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1360 }
1361
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]1362}
1363
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1364void
1365X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1366 MemOp2RegOpTableType &M2RTable,
1367 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1368 if ((Flags & TB_NO_FORWARD) == 0) {
1369 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1370 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1371 }
1372 if ((Flags & TB_NO_REVERSE) == 0) {
1373 assert(!M2RTable.count(MemOp) &&
1374 "Duplicated entries in unfolding maps?");
1375 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1376 }
1377}
1378
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1379bool
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1380X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1381 unsigned &SrcReg, unsigned &DstReg,
1382 unsigned &SubIdx) const {
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1383 switch (MI.getOpcode()) {
1384 default: break;
1385 case X86::MOVSX16rr8:
1386 case X86::MOVZX16rr8:
1387 case X86::MOVSX32rr8:
1388 case X86::MOVZX32rr8:
1389 case X86::MOVSX64rr8:
Evan Chengceb5a4e2010-01-13 08:01:32 +0000[diff] [blame]1390 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1391 // It's not always legal to reference the low 8-bit of the larger
1392 // register in 32-bit mode.
1393 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1394 case X86::MOVSX32rr16:
1395 case X86::MOVZX32rr16:
1396 case X86::MOVSX64rr16:
Tim Northover04eb4232013-05-30 10:43:18 +0000[diff] [blame]1397 case X86::MOVSX64rr32: {
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1398 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1399 // Be conservative.
1400 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1401 SrcReg = MI.getOperand(1).getReg();
1402 DstReg = MI.getOperand(0).getReg();
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1403 switch (MI.getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1404 default: llvm_unreachable("Unreachable!");
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1405 case X86::MOVSX16rr8:
1406 case X86::MOVZX16rr8:
1407 case X86::MOVSX32rr8:
1408 case X86::MOVZX32rr8:
1409 case X86::MOVSX64rr8:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1410 SubIdx = X86::sub_8bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1411 break;
1412 case X86::MOVSX32rr16:
1413 case X86::MOVZX32rr16:
1414 case X86::MOVSX64rr16:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1415 SubIdx = X86::sub_16bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1416 break;
1417 case X86::MOVSX64rr32:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1418 SubIdx = X86::sub_32bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1419 break;
1420 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1421 return true;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1422 }
1423 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1424 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1425}
1426
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1427/// isFrameOperand - Return true and the FrameIndex if the specified
1428/// operand and follow operands form a reference to the stack frame.
1429bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1430 int &FrameIndex) const {
1431 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1432 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1433 MI->getOperand(Op+1).getImm() == 1 &&
1434 MI->getOperand(Op+2).getReg() == 0 &&
1435 MI->getOperand(Op+3).getImm() == 0) {
1436 FrameIndex = MI->getOperand(Op).getIndex();
1437 return true;
1438 }
1439 return false;
1440}
1441
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1442static bool isFrameLoadOpcode(int Opcode) {
1443 switch (Opcode) {
David Blaikie46a9f012012-01-20 21:51:11 +0000[diff] [blame]1444 default:
1445 return false;
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1446 case X86::MOV8rm:
1447 case X86::MOV16rm:
1448 case X86::MOV32rm:
Evan Cheng11b0a5d2006-09-08 06:48:29 +0000[diff] [blame]1449 case X86::MOV64rm:
Dale Johannesen3d7008c2007-07-04 21:07:47 +0000[diff] [blame]1450 case X86::LD_Fp64m:
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1451 case X86::MOVSSrm:
1452 case X86::MOVSDrm:
Chris Lattnerbfc2c682006-04-18 16:44:51 +0000[diff] [blame]1453 case X86::MOVAPSrm:
1454 case X86::MOVAPDrm:
Dan Gohmanbdc0f8b2009-01-09 02:40:34 +0000[diff] [blame]1455 case X86::MOVDQArm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1456 case X86::VMOVSSrm:
1457 case X86::VMOVSDrm:
1458 case X86::VMOVAPSrm:
1459 case X86::VMOVAPDrm:
1460 case X86::VMOVDQArm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1461 case X86::VMOVAPSYrm:
1462 case X86::VMOVAPDYrm:
1463 case X86::VMOVDQAYrm:
Bill Wendlinge7b2a862007-04-03 06:00:37 +0000[diff] [blame]1464 case X86::MMX_MOVD64rm:
1465 case X86::MMX_MOVQ64rm:
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]1466 case X86::VMOVDQA32rm:
1467 case X86::VMOVDQA64rm:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1468 return true;
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1469 }
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1470}
1471
1472static bool isFrameStoreOpcode(int Opcode) {
1473 switch (Opcode) {
1474 default: break;
1475 case X86::MOV8mr:
1476 case X86::MOV16mr:
1477 case X86::MOV32mr:
1478 case X86::MOV64mr:
1479 case X86::ST_FpP64m:
1480 case X86::MOVSSmr:
1481 case X86::MOVSDmr:
1482 case X86::MOVAPSmr:
1483 case X86::MOVAPDmr:
1484 case X86::MOVDQAmr:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1485 case X86::VMOVSSmr:
1486 case X86::VMOVSDmr:
1487 case X86::VMOVAPSmr:
1488 case X86::VMOVAPDmr:
1489 case X86::VMOVDQAmr:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1490 case X86::VMOVAPSYmr:
1491 case X86::VMOVAPDYmr:
1492 case X86::VMOVDQAYmr:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1493 case X86::MMX_MOVD64mr:
1494 case X86::MMX_MOVQ64mr:
1495 case X86::MMX_MOVNTQmr:
1496 return true;
1497 }
1498 return false;
1499}
1500
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1501unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1502 int &FrameIndex) const {
1503 if (isFrameLoadOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1504 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1505 return MI->getOperand(0).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1506 return 0;
1507}
1508
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1509unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1510 int &FrameIndex) const {
1511 if (isFrameLoadOpcode(MI->getOpcode())) {
1512 unsigned Reg;
1513 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1514 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1515 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1516 const MachineMemOperand *Dummy;
1517 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1518 }
1519 return 0;
1520}
1521
Dan Gohman0b273252008-11-18 19:49:32 +0000[diff] [blame]1522unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1523 int &FrameIndex) const {
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1524 if (isFrameStoreOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1525 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1526 isFrameOperand(MI, 0, FrameIndex))
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]1527 return MI->getOperand(X86::AddrNumOperands).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1528 return 0;
1529}
1530
1531unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1532 int &FrameIndex) const {
1533 if (isFrameStoreOpcode(MI->getOpcode())) {
1534 unsigned Reg;
1535 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1536 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1537 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1538 const MachineMemOperand *Dummy;
1539 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1540 }
1541 return 0;
1542}
1543
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1544/// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1545/// X86::MOVPC32r.
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1546static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
Jakob Stoklund Olesen3b9a4422012-08-08 00:40:47 +0000[diff] [blame]1547 // Don't waste compile time scanning use-def chains of physregs.
1548 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1549 return false;
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1550 bool isPICBase = false;
1551 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1552 E = MRI.def_end(); I != E; ++I) {
1553 MachineInstr *DefMI = I.getOperand().getParent();
1554 if (DefMI->getOpcode() != X86::MOVPC32r)
1555 return false;
1556 assert(!isPICBase && "More than one PIC base?");
1557 isPICBase = true;
1558 }
1559 return isPICBase;
1560}
Evan Cheng1973a462008-03-31 07:54:19 +0000[diff] [blame]1561
Bill Wendling1e117682008-05-12 20:54:26 +0000[diff] [blame]1562bool
Dan Gohmane919de52009-10-10 00:34:18 +0000[diff] [blame]1563X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1564 AliasAnalysis *AA) const {
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1565 switch (MI->getOpcode()) {
1566 default: break;
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1567 case X86::MOV8rm:
1568 case X86::MOV16rm:
1569 case X86::MOV32rm:
1570 case X86::MOV64rm:
1571 case X86::LD_Fp64m:
1572 case X86::MOVSSrm:
1573 case X86::MOVSDrm:
1574 case X86::MOVAPSrm:
1575 case X86::MOVUPSrm:
1576 case X86::MOVAPDrm:
1577 case X86::MOVDQArm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1578 case X86::MOVDQUrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1579 case X86::VMOVSSrm:
1580 case X86::VMOVSDrm:
1581 case X86::VMOVAPSrm:
1582 case X86::VMOVUPSrm:
1583 case X86::VMOVAPDrm:
1584 case X86::VMOVDQArm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1585 case X86::VMOVDQUrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1586 case X86::VMOVAPSYrm:
1587 case X86::VMOVUPSYrm:
1588 case X86::VMOVAPDYrm:
1589 case X86::VMOVDQAYrm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1590 case X86::VMOVDQUYrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1591 case X86::MMX_MOVD64rm:
1592 case X86::MMX_MOVQ64rm:
1593 case X86::FsVMOVAPSrm:
1594 case X86::FsVMOVAPDrm:
1595 case X86::FsMOVAPSrm:
1596 case X86::FsMOVAPDrm: {
1597 // Loads from constant pools are trivially rematerializable.
1598 if (MI->getOperand(1).isReg() &&
1599 MI->getOperand(2).isImm() &&
1600 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1601 MI->isInvariantLoad(AA)) {
1602 unsigned BaseReg = MI->getOperand(1).getReg();
1603 if (BaseReg == 0 || BaseReg == X86::RIP)
1604 return true;
1605 // Allow re-materialization of PIC load.
1606 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1607 return false;
1608 const MachineFunction &MF = *MI->getParent()->getParent();
1609 const MachineRegisterInfo &MRI = MF.getRegInfo();
1610 return regIsPICBase(BaseReg, MRI);
Evan Cheng94ba37f2008-02-22 09:25:47 +0000[diff] [blame]1611 }
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1612 return false;
1613 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1614
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1615 case X86::LEA32r:
1616 case X86::LEA64r: {
1617 if (MI->getOperand(2).isImm() &&
1618 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1619 !MI->getOperand(4).isReg()) {
1620 // lea fi#, lea GV, etc. are all rematerializable.
1621 if (!MI->getOperand(1).isReg())
1622 return true;
1623 unsigned BaseReg = MI->getOperand(1).getReg();
1624 if (BaseReg == 0)
1625 return true;
1626 // Allow re-materialization of lea PICBase + x.
1627 const MachineFunction &MF = *MI->getParent()->getParent();
1628 const MachineRegisterInfo &MRI = MF.getRegInfo();
1629 return regIsPICBase(BaseReg, MRI);
1630 }
1631 return false;
1632 }
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1633 }
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1634
Dan Gohmane8c1e422007-06-26 00:48:07 +0000[diff] [blame]1635 // All other instructions marked M_REMATERIALIZABLE are always trivially
1636 // rematerializable.
1637 return true;
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1638}
1639
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1640/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1641/// would clobber the EFLAGS condition register. Note the result may be
1642/// conservative. If it cannot definitely determine the safety after visiting
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1643/// a few instructions in each direction it assumes it's not safe.
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1644static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1645 MachineBasicBlock::iterator I) {
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1646 MachineBasicBlock::iterator E = MBB.end();
1647
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1648 // For compile time consideration, if we are not able to determine the
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1649 // safety after visiting 4 instructions in each direction, we will assume
1650 // it's not safe.
1651 MachineBasicBlock::iterator Iter = I;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1652 for (unsigned i = 0; Iter != E && i < 4; ++i) {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1653 bool SeenDef = false;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1654 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1655 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1656 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1657 SeenDef = true;
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1658 if (!MO.isReg())
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1659 continue;
1660 if (MO.getReg() == X86::EFLAGS) {
1661 if (MO.isUse())
1662 return false;
1663 SeenDef = true;
1664 }
1665 }
1666
1667 if (SeenDef)
1668 // This instruction defines EFLAGS, no need to look any further.
1669 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1670 ++Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1671 // Skip over DBG_VALUE.
1672 while (Iter != E && Iter->isDebugValue())
1673 ++Iter;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1674 }
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1675
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1676 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1677 // live in.
1678 if (Iter == E) {
1679 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1680 SE = MBB.succ_end(); SI != SE; ++SI)
1681 if ((*SI)->isLiveIn(X86::EFLAGS))
1682 return false;
1683 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1684 }
1685
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1686 MachineBasicBlock::iterator B = MBB.begin();
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1687 Iter = I;
1688 for (unsigned i = 0; i < 4; ++i) {
1689 // 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]1690 // EFLAGS iff EFLAGS is not live-in.
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1691 if (Iter == B)
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1692 return !MBB.isLiveIn(X86::EFLAGS);
1693
1694 --Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1695 // Skip over DBG_VALUE.
1696 while (Iter != B && Iter->isDebugValue())
1697 --Iter;
1698
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1699 bool SawKill = false;
1700 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1701 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1702 // A register mask may clobber EFLAGS, but we should still look for a
1703 // live EFLAGS def.
1704 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1705 SawKill = true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1706 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1707 if (MO.isDef()) return MO.isDead();
1708 if (MO.isKill()) SawKill = true;
1709 }
1710 }
1711
1712 if (SawKill)
1713 // This instruction kills EFLAGS and doesn't redefine it, so
1714 // there's no need to look further.
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1715 return true;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1716 }
1717
1718 // Conservative answer.
1719 return false;
1720}
1721
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1722void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1723 MachineBasicBlock::iterator I,
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1724 unsigned DestReg, unsigned SubIdx,
Evan Cheng6ad7da92009-11-14 02:55:43 +0000[diff] [blame]1725 const MachineInstr *Orig,
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1726 const TargetRegisterInfo &TRI) const {
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]1727 // MOV32r0 is implemented with a xor which clobbers condition code.
1728 // Re-materialize it as movri instructions to avoid side effects.
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1729 unsigned Opc = Orig->getOpcode();
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]1730 if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
1731 DebugLoc DL = Orig->getDebugLoc();
1732 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
1733 .addImm(0);
1734 } else {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1735 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1736 MBB.insert(I, MI);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1737 }
Evan Cheng147cb762008-04-16 23:44:44 +0000[diff] [blame]1738
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1739 MachineInstr *NewMI = prior(I);
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1740 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1741}
1742
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1743/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1744/// is not marked dead.
1745static bool hasLiveCondCodeDef(MachineInstr *MI) {
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1746 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1747 MachineOperand &MO = MI->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1748 if (MO.isReg() && MO.isDef() &&
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1749 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1750 return true;
1751 }
1752 }
1753 return false;
1754}
1755
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]1756/// getTruncatedShiftCount - check whether the shift count for a machine operand
1757/// is non-zero.
1758inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
1759 unsigned ShiftAmtOperandIdx) {
1760 // The shift count is six bits with the REX.W prefix and five bits without.
1761 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
1762 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
1763 return Imm & ShiftCountMask;
1764}
1765
1766/// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
1767/// can be represented by a LEA instruction.
1768inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
1769 // Left shift instructions can be transformed into load-effective-address
1770 // instructions if we can encode them appropriately.
1771 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
1772 // The SIB.scale field is two bits wide which means that we can encode any
1773 // shift amount less than 4.
1774 return ShAmt < 4 && ShAmt > 0;
1775}
1776
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]1777bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
1778 unsigned Opc, bool AllowSP,
1779 unsigned &NewSrc, bool &isKill, bool &isUndef,
1780 MachineOperand &ImplicitOp) const {
1781 MachineFunction &MF = *MI->getParent()->getParent();
1782 const TargetRegisterClass *RC;
1783 if (AllowSP) {
1784 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
1785 } else {
1786 RC = Opc != X86::LEA32r ?
1787 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
1788 }
1789 unsigned SrcReg = Src.getReg();
1790
1791 // For both LEA64 and LEA32 the register already has essentially the right
1792 // type (32-bit or 64-bit) we may just need to forbid SP.
1793 if (Opc != X86::LEA64_32r) {
1794 NewSrc = SrcReg;
1795 isKill = Src.isKill();
1796 isUndef = Src.isUndef();
1797
1798 if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
1799 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
1800 return false;
1801
1802 return true;
1803 }
1804
1805 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
1806 // another we need to add 64-bit registers to the final MI.
1807 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
1808 ImplicitOp = Src;
1809 ImplicitOp.setImplicit();
1810
1811 NewSrc = getX86SubSuperRegister(Src.getReg(), MVT::i64);
1812 MachineBasicBlock::LivenessQueryResult LQR =
1813 MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
1814
1815 switch (LQR) {
1816 case MachineBasicBlock::LQR_Unknown:
1817 // We can't give sane liveness flags to the instruction, abandon LEA
1818 // formation.
1819 return false;
1820 case MachineBasicBlock::LQR_Live:
1821 isKill = MI->killsRegister(SrcReg);
1822 isUndef = false;
1823 break;
1824 default:
1825 // The physreg itself is dead, so we have to use it as an <undef>.
1826 isKill = false;
1827 isUndef = true;
1828 break;
1829 }
1830 } else {
1831 // Virtual register of the wrong class, we have to create a temporary 64-bit
1832 // vreg to feed into the LEA.
1833 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
1834 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1835 get(TargetOpcode::COPY))
1836 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
1837 .addOperand(Src);
1838
1839 // Which is obviously going to be dead after we're done with it.
1840 isKill = true;
1841 isUndef = false;
1842 }
1843
1844 // We've set all the parameters without issue.
1845 return true;
1846}
1847
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1848/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1849/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1850/// to a 32-bit superregister and then truncating back down to a 16-bit
1851/// subregister.
1852MachineInstr *
1853X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1854 MachineFunction::iterator &MFI,
1855 MachineBasicBlock::iterator &MBBI,
1856 LiveVariables *LV) const {
1857 MachineInstr *MI = MBBI;
1858 unsigned Dest = MI->getOperand(0).getReg();
1859 unsigned Src = MI->getOperand(1).getReg();
1860 bool isDead = MI->getOperand(0).isDead();
1861 bool isKill = MI->getOperand(1).isKill();
1862
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1863 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1864 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]1865 unsigned Opc, leaInReg;
1866 if (TM.getSubtarget<X86Subtarget>().is64Bit()) {
1867 Opc = X86::LEA64_32r;
1868 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1869 } else {
1870 Opc = X86::LEA32r;
1871 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1872 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1873
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1874 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1875 // well be shifting and then extracting the lower 16-bits.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1876 // This has the potential to cause partial register stall. e.g.
Evan Cheng3974c8d2009-12-12 18:55:26 +0000[diff] [blame]1877 // movw (%rbp,%rcx,2), %dx
1878 // leal -65(%rdx), %esi
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1879 // But testing has shown this *does* help performance in 64-bit mode (at
1880 // least on modern x86 machines).
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1881 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1882 MachineInstr *InsMI =
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1883 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1884 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1885 .addReg(Src, getKillRegState(isKill));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1886
1887 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1888 get(Opc), leaOutReg);
1889 switch (MIOpc) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1890 default: llvm_unreachable("Unreachable!");
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1891 case X86::SHL16ri: {
1892 unsigned ShAmt = MI->getOperand(2).getImm();
1893 MIB.addReg(0).addImm(1 << ShAmt)
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1894 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1895 break;
1896 }
1897 case X86::INC16r:
1898 case X86::INC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1899 addRegOffset(MIB, leaInReg, true, 1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1900 break;
1901 case X86::DEC16r:
1902 case X86::DEC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1903 addRegOffset(MIB, leaInReg, true, -1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1904 break;
1905 case X86::ADD16ri:
1906 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]1907 case X86::ADD16ri_DB:
1908 case X86::ADD16ri8_DB:
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1909 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1910 break;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1911 case X86::ADD16rr:
1912 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1913 unsigned Src2 = MI->getOperand(2).getReg();
1914 bool isKill2 = MI->getOperand(2).isKill();
1915 unsigned leaInReg2 = 0;
1916 MachineInstr *InsMI2 = 0;
1917 if (Src == Src2) {
1918 // ADD16rr %reg1028<kill>, %reg1028
1919 // just a single insert_subreg.
1920 addRegReg(MIB, leaInReg, true, leaInReg, false);
1921 } else {
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]1922 if (TM.getSubtarget<X86Subtarget>().is64Bit())
1923 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1924 else
1925 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1926 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1927 // well be shifting and then extracting the lower 16-bits.
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1928 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1929 InsMI2 =
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1930 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1931 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1932 .addReg(Src2, getKillRegState(isKill2));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1933 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1934 }
1935 if (LV && isKill2 && InsMI2)
1936 LV->replaceKillInstruction(Src2, MI, InsMI2);
1937 break;
1938 }
1939 }
1940
1941 MachineInstr *NewMI = MIB;
1942 MachineInstr *ExtMI =
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1943 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1944 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1945 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1946
1947 if (LV) {
1948 // Update live variables
1949 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1950 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1951 if (isKill)
1952 LV->replaceKillInstruction(Src, MI, InsMI);
1953 if (isDead)
1954 LV->replaceKillInstruction(Dest, MI, ExtMI);
1955 }
1956
1957 return ExtMI;
1958}
1959
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1960/// convertToThreeAddress - This method must be implemented by targets that
1961/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1962/// may be able to convert a two-address instruction into a true
1963/// three-address instruction on demand. This allows the X86 target (for
1964/// example) to convert ADD and SHL instructions into LEA instructions if they
1965/// would require register copies due to two-addressness.
1966///
1967/// This method returns a null pointer if the transformation cannot be
1968/// performed, otherwise it returns the new instruction.
1969///
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1970MachineInstr *
1971X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1972 MachineBasicBlock::iterator &MBBI,
Owen Anderson30cc0282008-07-02 23:41:07 +0000[diff] [blame]1973 LiveVariables *LV) const {
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1974 MachineInstr *MI = MBBI;
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]1975
1976 // The following opcodes also sets the condition code register(s). Only
1977 // convert them to equivalent lea if the condition code register def's
1978 // are dead!
1979 if (hasLiveCondCodeDef(MI))
1980 return 0;
1981
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1982 MachineFunction &MF = *MI->getParent()->getParent();
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1983 // All instructions input are two-addr instructions. Get the known operands.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1984 const MachineOperand &Dest = MI->getOperand(0);
1985 const MachineOperand &Src = MI->getOperand(1);
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1986
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]1987 MachineInstr *NewMI = NULL;
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1988 // 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]1989 // we have better subtarget support, enable the 16-bit LEA generation here.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1990 // 16-bit LEA is also slow on Core2.
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1991 bool DisableLEA16 = true;
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1992 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1993
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1994 unsigned MIOpc = MI->getOpcode();
1995 switch (MIOpc) {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1996 case X86::SHUFPSrri: {
1997 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1998 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1999
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]2000 unsigned B = MI->getOperand(1).getReg();
2001 unsigned C = MI->getOperand(2).getReg();
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2002 if (B != C) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2003 unsigned M = MI->getOperand(3).getImm();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2004 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2005 .addOperand(Dest).addOperand(Src).addImm(M);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2006 break;
2007 }
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]2008 case X86::SHUFPDrri: {
2009 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
2010 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
2011
2012 unsigned B = MI->getOperand(1).getReg();
2013 unsigned C = MI->getOperand(2).getReg();
2014 if (B != C) return 0;
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]2015 unsigned M = MI->getOperand(3).getImm();
2016
2017 // Convert to PSHUFD mask.
2018 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
2019
2020 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2021 .addOperand(Dest).addOperand(Src).addImm(M);
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]2022 break;
2023 }
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]2024 case X86::SHL64ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]2025 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]2026 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2027 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2028
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2029 // LEA can't handle RSP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2030 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2031 !MF.getRegInfo().constrainRegClass(Src.getReg(),
2032 &X86::GR64_NOSPRegClass))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2033 return 0;
2034
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2035 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2036 .addOperand(Dest)
2037 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]2038 break;
2039 }
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2040 case X86::SHL32ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]2041 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]2042 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2043 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2044
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2045 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2046
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2047 // LEA can't handle ESP.
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2048 bool isKill, isUndef;
2049 unsigned SrcReg;
2050 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2051 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2052 SrcReg, isKill, isUndef, ImplicitOp))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2053 return 0;
2054
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2055 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2056 .addOperand(Dest)
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2057 .addReg(0).addImm(1 << ShAmt)
2058 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2059 .addImm(0).addReg(0);
2060 if (ImplicitOp.getReg() != 0)
2061 MIB.addOperand(ImplicitOp);
2062 NewMI = MIB;
2063
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2064 break;
2065 }
2066 case X86::SHL16ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]2067 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]2068 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2069 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2070
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2071 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2072 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2073 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2074 .addOperand(Dest)
2075 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2076 break;
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]2077 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2078 default: {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]2079
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2080 switch (MIOpc) {
2081 default: return 0;
2082 case X86::INC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]2083 case X86::INC32r:
2084 case X86::INC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2085 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]2086 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2087 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2088 bool isKill, isUndef;
2089 unsigned SrcReg;
2090 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2091 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2092 SrcReg, isKill, isUndef, ImplicitOp))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2093 return 0;
2094
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2095 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2096 .addOperand(Dest)
2097 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef));
2098 if (ImplicitOp.getReg() != 0)
2099 MIB.addOperand(ImplicitOp);
2100
2101 NewMI = addOffset(MIB, 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2102 break;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2103 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2104 case X86::INC16r:
2105 case X86::INC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2106 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2107 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2108 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2109 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2110 .addOperand(Dest).addOperand(Src), 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2111 break;
2112 case X86::DEC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]2113 case X86::DEC32r:
2114 case X86::DEC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2115 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]2116 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2117 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2118
2119 bool isKill, isUndef;
2120 unsigned SrcReg;
2121 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2122 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2123 SrcReg, isKill, isUndef, ImplicitOp))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2124 return 0;
2125
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2126 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2127 .addOperand(Dest)
2128 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2129 if (ImplicitOp.getReg() != 0)
2130 MIB.addOperand(ImplicitOp);
2131
2132 NewMI = addOffset(MIB, -1);
2133
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2134 break;
2135 }
2136 case X86::DEC16r:
2137 case X86::DEC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2138 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2139 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2140 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2141 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2142 .addOperand(Dest).addOperand(Src), -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2143 break;
2144 case X86::ADD64rr:
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2145 case X86::ADD64rr_DB:
2146 case X86::ADD32rr:
2147 case X86::ADD32rr_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2148 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2149 unsigned Opc;
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2150 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2151 Opc = X86::LEA64r;
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2152 else
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2153 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2154
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2155 bool isKill, isUndef;
2156 unsigned SrcReg;
2157 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2158 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2159 SrcReg, isKill, isUndef, ImplicitOp))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2160 return 0;
2161
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2162 const MachineOperand &Src2 = MI->getOperand(2);
2163 bool isKill2, isUndef2;
2164 unsigned SrcReg2;
2165 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2166 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2167 SrcReg2, isKill2, isUndef2, ImplicitOp2))
2168 return 0;
2169
2170 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2171 .addOperand(Dest);
2172 if (ImplicitOp.getReg() != 0)
2173 MIB.addOperand(ImplicitOp);
2174 if (ImplicitOp2.getReg() != 0)
2175 MIB.addOperand(ImplicitOp2);
2176
2177 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
Nadav Rotem4968e452012-07-16 10:52:25 +0000[diff] [blame]2178
2179 // Preserve undefness of the operands.
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2180 NewMI->getOperand(1).setIsUndef(isUndef);
2181 NewMI->getOperand(3).setIsUndef(isUndef2);
Nadav Rotem4968e452012-07-16 10:52:25 +0000[diff] [blame]2182
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2183 if (LV && Src2.isKill())
2184 LV->replaceKillInstruction(SrcReg2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2185 break;
2186 }
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2187 case X86::ADD16rr:
2188 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2189 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2190 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2191 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2192 unsigned Src2 = MI->getOperand(2).getReg();
2193 bool isKill2 = MI->getOperand(2).isKill();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2194 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2195 .addOperand(Dest),
2196 Src.getReg(), Src.isKill(), Src2, isKill2);
2197
2198 // Preserve undefness of the operands.
2199 bool isUndef = MI->getOperand(1).isUndef();
2200 bool isUndef2 = MI->getOperand(2).isUndef();
2201 NewMI->getOperand(1).setIsUndef(isUndef);
2202 NewMI->getOperand(3).setIsUndef(isUndef2);
2203
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2204 if (LV && isKill2)
2205 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2206 break;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2207 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2208 case X86::ADD64ri32:
2209 case X86::ADD64ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2210 case X86::ADD64ri32_DB:
2211 case X86::ADD64ri8_DB:
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2212 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2213 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2214 .addOperand(Dest).addOperand(Src),
2215 MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2216 break;
2217 case X86::ADD32ri:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2218 case X86::ADD32ri8:
2219 case X86::ADD32ri_DB:
2220 case X86::ADD32ri8_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2221 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2222 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Tim Northover6833e3f2013-06-10 20:43:49 +0000[diff] [blame]2223
2224 bool isKill, isUndef;
2225 unsigned SrcReg;
2226 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2227 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2228 SrcReg, isKill, isUndef, ImplicitOp))
2229 return 0;
2230
2231 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2232 .addOperand(Dest)
2233 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2234 if (ImplicitOp.getReg() != 0)
2235 MIB.addOperand(ImplicitOp);
2236
2237 NewMI = addOffset(MIB, MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2238 break;
2239 }
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2240 case X86::ADD16ri:
2241 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2242 case X86::ADD16ri_DB:
2243 case X86::ADD16ri8_DB:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2244 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2245 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2246 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2247 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2248 .addOperand(Dest).addOperand(Src),
2249 MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2250 break;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2251 }
2252 }
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2253 }
2254
Evan Cheng1bc1cae2008-02-07 08:29:53 +0000[diff] [blame]2255 if (!NewMI) return 0;
2256
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2257 if (LV) { // Update live variables
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2258 if (Src.isKill())
2259 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2260 if (Dest.isDead())
2261 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2262 }
2263
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2264 MFI->insert(MBBI, NewMI); // Insert the new inst
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]2265 return NewMI;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2266}
2267
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2268/// commuteInstruction - We have a few instructions that must be hacked on to
2269/// commute them.
2270///
Evan Cheng03553bb2008-06-16 07:33:11 +0000[diff] [blame]2271MachineInstr *
2272X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2273 switch (MI->getOpcode()) {
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2274 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2275 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2276 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2277 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2278 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2279 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2280 unsigned Opc;
2281 unsigned Size;
2282 switch (MI->getOpcode()) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2283 default: llvm_unreachable("Unreachable!");
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2284 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2285 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2286 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2287 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2288 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2289 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2290 }
Chris Lattner5c463782007-12-30 20:49:49 +0000[diff] [blame]2291 unsigned Amt = MI->getOperand(3).getImm();
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2292 if (NewMI) {
2293 MachineFunction &MF = *MI->getParent()->getParent();
2294 MI = MF.CloneMachineInstr(MI);
2295 NewMI = false;
Evan Cheng244183e2008-02-13 02:46:49 +0000[diff] [blame]2296 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2297 MI->setDesc(get(Opc));
2298 MI->getOperand(3).setImm(Size-Amt);
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]2299 return TargetInstrInfo::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2300 }
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2301 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2302 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2303 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2304 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2305 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2306 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2307 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2308 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2309 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2310 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2311 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2312 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2313 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2314 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2315 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2316 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2317 unsigned Opc;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2318 switch (MI->getOpcode()) {
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2319 default: llvm_unreachable("Unreachable!");
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2320 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2321 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2322 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2323 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2324 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2325 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2326 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2327 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2328 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2329 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2330 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2331 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
Chris Lattner1a1c6002010-10-05 23:00:14 +0000[diff] [blame]2332 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2333 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2334 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2335 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2336 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2337 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2338 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2339 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2340 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2341 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2342 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2343 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2344 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2345 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2346 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2347 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2348 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2349 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2350 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2351 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2352 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2353 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2354 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2355 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2356 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2357 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2358 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2359 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2360 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2361 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2362 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2363 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2364 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2365 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2366 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2367 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2368 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2369 if (NewMI) {
2370 MachineFunction &MF = *MI->getParent()->getParent();
2371 MI = MF.CloneMachineInstr(MI);
2372 NewMI = false;
2373 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]2374 MI->setDesc(get(Opc));
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2375 // Fallthrough intended.
2376 }
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2377 default:
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]2378 return TargetInstrInfo::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2379 }
2380}
2381
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2382static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2383 switch (BrOpc) {
2384 default: return X86::COND_INVALID;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2385 case X86::JE_4: return X86::COND_E;
2386 case X86::JNE_4: return X86::COND_NE;
2387 case X86::JL_4: return X86::COND_L;
2388 case X86::JLE_4: return X86::COND_LE;
2389 case X86::JG_4: return X86::COND_G;
2390 case X86::JGE_4: return X86::COND_GE;
2391 case X86::JB_4: return X86::COND_B;
2392 case X86::JBE_4: return X86::COND_BE;
2393 case X86::JA_4: return X86::COND_A;
2394 case X86::JAE_4: return X86::COND_AE;
2395 case X86::JS_4: return X86::COND_S;
2396 case X86::JNS_4: return X86::COND_NS;
2397 case X86::JP_4: return X86::COND_P;
2398 case X86::JNP_4: return X86::COND_NP;
2399 case X86::JO_4: return X86::COND_O;
2400 case X86::JNO_4: return X86::COND_NO;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2401 }
2402}
2403
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2404/// getCondFromSETOpc - return condition code of a SET opcode.
2405static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2406 switch (Opc) {
2407 default: return X86::COND_INVALID;
2408 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2409 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2410 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2411 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2412 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2413 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2414 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2415 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2416 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2417 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2418 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2419 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2420 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2421 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2422 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2423 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2424 }
2425}
2426
2427/// getCondFromCmovOpc - return condition code of a CMov opcode.
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]2428X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2429 switch (Opc) {
2430 default: return X86::COND_INVALID;
2431 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2432 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2433 return X86::COND_A;
2434 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2435 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2436 return X86::COND_AE;
2437 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2438 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2439 return X86::COND_B;
2440 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2441 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2442 return X86::COND_BE;
2443 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2444 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2445 return X86::COND_E;
2446 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2447 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2448 return X86::COND_G;
2449 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2450 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2451 return X86::COND_GE;
2452 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2453 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2454 return X86::COND_L;
2455 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2456 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2457 return X86::COND_LE;
2458 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2459 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2460 return X86::COND_NE;
2461 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2462 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2463 return X86::COND_NO;
2464 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2465 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2466 return X86::COND_NP;
2467 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2468 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2469 return X86::COND_NS;
2470 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2471 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2472 return X86::COND_O;
2473 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2474 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2475 return X86::COND_P;
2476 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2477 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2478 return X86::COND_S;
2479 }
2480}
2481
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2482unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2483 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2484 default: llvm_unreachable("Illegal condition code!");
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2485 case X86::COND_E: return X86::JE_4;
2486 case X86::COND_NE: return X86::JNE_4;
2487 case X86::COND_L: return X86::JL_4;
2488 case X86::COND_LE: return X86::JLE_4;
2489 case X86::COND_G: return X86::JG_4;
2490 case X86::COND_GE: return X86::JGE_4;
2491 case X86::COND_B: return X86::JB_4;
2492 case X86::COND_BE: return X86::JBE_4;
2493 case X86::COND_A: return X86::JA_4;
2494 case X86::COND_AE: return X86::JAE_4;
2495 case X86::COND_S: return X86::JS_4;
2496 case X86::COND_NS: return X86::JNS_4;
2497 case X86::COND_P: return X86::JP_4;
2498 case X86::COND_NP: return X86::JNP_4;
2499 case X86::COND_O: return X86::JO_4;
2500 case X86::COND_NO: return X86::JNO_4;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2501 }
2502}
2503
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2504/// GetOppositeBranchCondition - Return the inverse of the specified condition,
2505/// e.g. turning COND_E to COND_NE.
2506X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2507 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2508 default: llvm_unreachable("Illegal condition code!");
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2509 case X86::COND_E: return X86::COND_NE;
2510 case X86::COND_NE: return X86::COND_E;
2511 case X86::COND_L: return X86::COND_GE;
2512 case X86::COND_LE: return X86::COND_G;
2513 case X86::COND_G: return X86::COND_LE;
2514 case X86::COND_GE: return X86::COND_L;
2515 case X86::COND_B: return X86::COND_AE;
2516 case X86::COND_BE: return X86::COND_A;
2517 case X86::COND_A: return X86::COND_BE;
2518 case X86::COND_AE: return X86::COND_B;
2519 case X86::COND_S: return X86::COND_NS;
2520 case X86::COND_NS: return X86::COND_S;
2521 case X86::COND_P: return X86::COND_NP;
2522 case X86::COND_NP: return X86::COND_P;
2523 case X86::COND_O: return X86::COND_NO;
2524 case X86::COND_NO: return X86::COND_O;
2525 }
2526}
2527
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2528/// getSwappedCondition - assume the flags are set by MI(a,b), return
2529/// the condition code if we modify the instructions such that flags are
2530/// set by MI(b,a).
Benjamin Kramerabbfe692012-07-13 13:25:15 +0000[diff] [blame]2531static X86::CondCode getSwappedCondition(X86::CondCode CC) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2532 switch (CC) {
2533 default: return X86::COND_INVALID;
2534 case X86::COND_E: return X86::COND_E;
2535 case X86::COND_NE: return X86::COND_NE;
2536 case X86::COND_L: return X86::COND_G;
2537 case X86::COND_LE: return X86::COND_GE;
2538 case X86::COND_G: return X86::COND_L;
2539 case X86::COND_GE: return X86::COND_LE;
2540 case X86::COND_B: return X86::COND_A;
2541 case X86::COND_BE: return X86::COND_AE;
2542 case X86::COND_A: return X86::COND_B;
2543 case X86::COND_AE: return X86::COND_BE;
2544 }
2545}
2546
2547/// getSETFromCond - Return a set opcode for the given condition and
2548/// whether it has memory operand.
2549static unsigned getSETFromCond(X86::CondCode CC,
2550 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2551 static const uint16_t Opc[16][2] = {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2552 { X86::SETAr, X86::SETAm },
2553 { X86::SETAEr, X86::SETAEm },
2554 { X86::SETBr, X86::SETBm },
2555 { X86::SETBEr, X86::SETBEm },
2556 { X86::SETEr, X86::SETEm },
2557 { X86::SETGr, X86::SETGm },
2558 { X86::SETGEr, X86::SETGEm },
2559 { X86::SETLr, X86::SETLm },
2560 { X86::SETLEr, X86::SETLEm },
2561 { X86::SETNEr, X86::SETNEm },
2562 { X86::SETNOr, X86::SETNOm },
2563 { X86::SETNPr, X86::SETNPm },
2564 { X86::SETNSr, X86::SETNSm },
2565 { X86::SETOr, X86::SETOm },
2566 { X86::SETPr, X86::SETPm },
2567 { X86::SETSr, X86::SETSm }
2568 };
2569
2570 assert(CC < 16 && "Can only handle standard cond codes");
2571 return Opc[CC][HasMemoryOperand ? 1 : 0];
2572}
2573
2574/// getCMovFromCond - Return a cmov opcode for the given condition,
2575/// register size in bytes, and operand type.
2576static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2577 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2578 static const uint16_t Opc[32][3] = {
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2579 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2580 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2581 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2582 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2583 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2584 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2585 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2586 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2587 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2588 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2589 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2590 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2591 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2592 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2593 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2594 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2595 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2596 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2597 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2598 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2599 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2600 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2601 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2602 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2603 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2604 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2605 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2606 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2607 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2608 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2609 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2610 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2611 };
2612
2613 assert(CC < 16 && "Can only handle standard cond codes");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2614 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2615 switch(RegBytes) {
2616 default: llvm_unreachable("Illegal register size!");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2617 case 2: return Opc[Idx][0];
2618 case 4: return Opc[Idx][1];
2619 case 8: return Opc[Idx][2];
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2620 }
2621}
2622
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2623bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2624 if (!MI->isTerminator()) return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2625
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2626 // Conditional branch is a special case.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2627 if (MI->isBranch() && !MI->isBarrier())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2628 return true;
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2629 if (!MI->isPredicable())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2630 return true;
2631 return !isPredicated(MI);
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2632}
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2633
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2634bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2635 MachineBasicBlock *&TBB,
2636 MachineBasicBlock *&FBB,
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2637 SmallVectorImpl<MachineOperand> &Cond,
2638 bool AllowModify) const {
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2639 // Start from the bottom of the block and work up, examining the
2640 // terminator instructions.
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2641 MachineBasicBlock::iterator I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2642 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2643 while (I != MBB.begin()) {
2644 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2645 if (I->isDebugValue())
2646 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2647
2648 // Working from the bottom, when we see a non-terminator instruction, we're
2649 // done.
Jakob Stoklund Olesenc30b4dd2010-07-16 17:41:44 +0000[diff] [blame]2650 if (!isUnpredicatedTerminator(I))
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2651 break;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2652
2653 // A terminator that isn't a branch can't easily be handled by this
2654 // analysis.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2655 if (!I->isBranch())
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2656 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2657
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2658 // Handle unconditional branches.
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2659 if (I->getOpcode() == X86::JMP_4) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2660 UnCondBrIter = I;
2661
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2662 if (!AllowModify) {
2663 TBB = I->getOperand(0).getMBB();
Evan Cheng2fa28112009-05-08 06:34:09 +0000[diff] [blame]2664 continue;
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2665 }
2666
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2667 // If the block has any instructions after a JMP, delete them.
Chris Lattnera48f44d2009-12-03 00:50:42 +0000[diff] [blame]2668 while (llvm::next(I) != MBB.end())
2669 llvm::next(I)->eraseFromParent();
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2670
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2671 Cond.clear();
2672 FBB = 0;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2673
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2674 // Delete the JMP if it's equivalent to a fall-through.
2675 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2676 TBB = 0;
2677 I->eraseFromParent();
2678 I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2679 UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2680 continue;
2681 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2682
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2683 // TBB is used to indicate the unconditional destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2684 TBB = I->getOperand(0).getMBB();
2685 continue;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2686 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2687
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2688 // Handle conditional branches.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2689 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2690 if (BranchCode == X86::COND_INVALID)
2691 return true; // Can't handle indirect branch.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2692
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2693 // Working from the bottom, handle the first conditional branch.
2694 if (Cond.empty()) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2695 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2696 if (AllowModify && UnCondBrIter != MBB.end() &&
2697 MBB.isLayoutSuccessor(TargetBB)) {
2698 // If we can modify the code and it ends in something like:
2699 //
2700 // jCC L1
2701 // jmp L2
2702 // L1:
2703 // ...
2704 // L2:
2705 //
2706 // Then we can change this to:
2707 //
2708 // jnCC L2
2709 // L1:
2710 // ...
2711 // L2:
2712 //
2713 // Which is a bit more efficient.
2714 // We conditionally jump to the fall-through block.
2715 BranchCode = GetOppositeBranchCondition(BranchCode);
2716 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2717 MachineBasicBlock::iterator OldInst = I;
2718
2719 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2720 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2721 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2722 .addMBB(TargetBB);
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2723
2724 OldInst->eraseFromParent();
2725 UnCondBrIter->eraseFromParent();
2726
2727 // Restart the analysis.
2728 UnCondBrIter = MBB.end();
2729 I = MBB.end();
2730 continue;
2731 }
2732
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2733 FBB = TBB;
2734 TBB = I->getOperand(0).getMBB();
2735 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2736 continue;
2737 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2738
2739 // Handle subsequent conditional branches. Only handle the case where all
2740 // conditional branches branch to the same destination and their condition
2741 // opcodes fit one of the special multi-branch idioms.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2742 assert(Cond.size() == 1);
2743 assert(TBB);
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2744
2745 // Only handle the case where all conditional branches branch to the same
2746 // destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2747 if (TBB != I->getOperand(0).getMBB())
2748 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2749
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2750 // If the conditions are the same, we can leave them alone.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2751 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2752 if (OldBranchCode == BranchCode)
2753 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2754
2755 // If they differ, see if they fit one of the known patterns. Theoretically,
2756 // we could handle more patterns here, but we shouldn't expect to see them
2757 // if instruction selection has done a reasonable job.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2758 if ((OldBranchCode == X86::COND_NP &&
2759 BranchCode == X86::COND_E) ||
2760 (OldBranchCode == X86::COND_E &&
2761 BranchCode == X86::COND_NP))
2762 BranchCode = X86::COND_NP_OR_E;
2763 else if ((OldBranchCode == X86::COND_P &&
2764 BranchCode == X86::COND_NE) ||
2765 (OldBranchCode == X86::COND_NE &&
2766 BranchCode == X86::COND_P))
2767 BranchCode = X86::COND_NE_OR_P;
2768 else
2769 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2770
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2771 // Update the MachineOperand.
2772 Cond[0].setImm(BranchCode);
Chris Lattner74436002006-10-30 22:27:23 +0000[diff] [blame]2773 }
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2774
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2775 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2776}
2777
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2778unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2779 MachineBasicBlock::iterator I = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2780 unsigned Count = 0;
2781
2782 while (I != MBB.begin()) {
2783 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2784 if (I->isDebugValue())
2785 continue;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2786 if (I->getOpcode() != X86::JMP_4 &&
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2787 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2788 break;
2789 // Remove the branch.
2790 I->eraseFromParent();
2791 I = MBB.end();
2792 ++Count;
2793 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2794
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2795 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2796}
2797
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2798unsigned
2799X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2800 MachineBasicBlock *FBB,
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2801 const SmallVectorImpl<MachineOperand> &Cond,
2802 DebugLoc DL) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2803 // Shouldn't be a fall through.
2804 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
Chris Lattner6fca75e2006-10-21 05:34:23 +0000[diff] [blame]2805 assert((Cond.size() == 1 || Cond.size() == 0) &&
2806 "X86 branch conditions have one component!");
2807
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2808 if (Cond.empty()) {
2809 // Unconditional branch?
2810 assert(!FBB && "Unconditional branch with multiple successors!");
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2811 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2812 return 1;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2813 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2814
2815 // Conditional branch.
2816 unsigned Count = 0;
2817 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2818 switch (CC) {
2819 case X86::COND_NP_OR_E:
2820 // Synthesize NP_OR_E with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2821 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2822 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2823 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2824 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2825 break;
2826 case X86::COND_NE_OR_P:
2827 // Synthesize NE_OR_P with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2828 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2829 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2830 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2831 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2832 break;
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2833 default: {
2834 unsigned Opc = GetCondBranchFromCond(CC);
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2835 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2836 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2837 }
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2838 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2839 if (FBB) {
2840 // Two-way Conditional branch. Insert the second branch.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2841 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2842 ++Count;
2843 }
2844 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2845}
2846
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2847bool X86InstrInfo::
2848canInsertSelect(const MachineBasicBlock &MBB,
2849 const SmallVectorImpl<MachineOperand> &Cond,
2850 unsigned TrueReg, unsigned FalseReg,
2851 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2852 // Not all subtargets have cmov instructions.
2853 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2854 return false;
2855 if (Cond.size() != 1)
2856 return false;
2857 // We cannot do the composite conditions, at least not in SSA form.
2858 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2859 return false;
2860
2861 // Check register classes.
2862 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2863 const TargetRegisterClass *RC =
2864 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2865 if (!RC)
2866 return false;
2867
2868 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2869 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2870 X86::GR32RegClass.hasSubClassEq(RC) ||
2871 X86::GR64RegClass.hasSubClassEq(RC)) {
2872 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2873 // Bridge. Probably Ivy Bridge as well.
2874 CondCycles = 2;
2875 TrueCycles = 2;
2876 FalseCycles = 2;
2877 return true;
2878 }
2879
2880 // Can't do vectors.
2881 return false;
2882}
2883
2884void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2885 MachineBasicBlock::iterator I, DebugLoc DL,
2886 unsigned DstReg,
2887 const SmallVectorImpl<MachineOperand> &Cond,
2888 unsigned TrueReg, unsigned FalseReg) const {
2889 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2890 assert(Cond.size() == 1 && "Invalid Cond array");
2891 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2892 MRI.getRegClass(DstReg)->getSize(),
2893 false/*HasMemoryOperand*/);
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2894 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2895}
2896
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2897/// isHReg - Test if the given register is a physical h register.
2898static bool isHReg(unsigned Reg) {
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2899 return X86::GR8_ABCD_HRegClass.contains(Reg);
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2900}
2901
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2902// Try and copy between VR128/VR64 and GR64 registers.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2903static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]2904 const X86Subtarget& Subtarget) {
2905
2906
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2907 // SrcReg(VR128) -> DestReg(GR64)
2908 // SrcReg(VR64) -> DestReg(GR64)
2909 // SrcReg(GR64) -> DestReg(VR128)
2910 // SrcReg(GR64) -> DestReg(VR64)
2911
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]2912 bool HasAVX = Subtarget.hasAVX();
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2913 bool HasAVX512 = Subtarget.hasAVX512();
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2914 if (X86::GR64RegClass.contains(DestReg)) {
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2915 if (X86::VR128XRegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2916 // Copy from a VR128 register to a GR64 register.
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2917 return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
2918 X86::MOVPQIto64rr);
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2919 if (X86::VR64RegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2920 // Copy from a VR64 register to a GR64 register.
2921 return X86::MOVSDto64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2922 } else if (X86::GR64RegClass.contains(SrcReg)) {
2923 // Copy from a GR64 register to a VR128 register.
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2924 if (X86::VR128XRegClass.contains(DestReg))
2925 return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
2926 X86::MOV64toPQIrr);
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2927 // Copy from a GR64 register to a VR64 register.
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2928 if (X86::VR64RegClass.contains(DestReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2929 return X86::MOV64toSDrr;
2930 }
2931
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2932 // SrcReg(FR32) -> DestReg(GR32)
2933 // SrcReg(GR32) -> DestReg(FR32)
2934
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2935 if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2936 // Copy from a FR32 register to a GR32 register.
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2937 return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2938
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2939 if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2940 // Copy from a GR32 register to a FR32 register.
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]2941 return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2942 return 0;
2943}
2944
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]2945static
2946unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg) {
2947 if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
2948 X86::VR256XRegClass.contains(DestReg, SrcReg) ||
2949 X86::VR512RegClass.contains(DestReg, SrcReg)) {
2950 DestReg = get512BitSuperRegister(DestReg);
2951 SrcReg = get512BitSuperRegister(SrcReg);
2952 return X86::VMOVAPSZrr;
2953 }
2954 if ((X86::VK8RegClass.contains(DestReg) ||
2955 X86::VK16RegClass.contains(DestReg)) &&
2956 (X86::VK8RegClass.contains(SrcReg) ||
2957 X86::VK16RegClass.contains(SrcReg)))
2958 return X86::KMOVWkk;
2959 return 0;
2960}
2961
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2962void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2963 MachineBasicBlock::iterator MI, DebugLoc DL,
2964 unsigned DestReg, unsigned SrcReg,
2965 bool KillSrc) const {
2966 // First deal with the normal symmetric copies.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2967 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]2968 bool HasAVX512 = TM.getSubtarget<X86Subtarget>().hasAVX512();
2969 unsigned Opc = 0;
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2970 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2971 Opc = X86::MOV64rr;
2972 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2973 Opc = X86::MOV32rr;
2974 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2975 Opc = X86::MOV16rr;
2976 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2977 // Copying to or from a physical H register on x86-64 requires a NOREX
2978 // move. Otherwise use a normal move.
2979 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2980 TM.getSubtarget<X86Subtarget>().is64Bit()) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2981 Opc = X86::MOV8rr_NOREX;
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2982 // Both operands must be encodable without an REX prefix.
2983 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2984 "8-bit H register can not be copied outside GR8_NOREX");
2985 } else
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2986 Opc = X86::MOV8rr;
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]2987 }
2988 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2989 Opc = X86::MMX_MOVQ64rr;
2990 else if (HasAVX512)
2991 Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg);
2992 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2993 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]2994 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2995 Opc = X86::VMOVAPSYrr;
Elena Demikhovskycf5b1452013-08-11 07:55:09 +0000[diff] [blame]2996 if (!Opc)
2997 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, TM.getSubtarget<X86Subtarget>());
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2998
2999 if (Opc) {
3000 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3001 .addReg(SrcReg, getKillRegState(KillSrc));
3002 return;
3003 }
3004
3005 // Moving EFLAGS to / from another register requires a push and a pop.
Nadav Rotemd5aae982012-12-21 23:48:49 +0000[diff] [blame]3006 // Notice that we have to adjust the stack if we don't want to clobber the
3007 // first frame index. See X86FrameLowering.cpp - colobbersTheStack.
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3008 if (SrcReg == X86::EFLAGS) {
3009 if (X86::GR64RegClass.contains(DestReg)) {
3010 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
3011 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
3012 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]3013 }
3014 if (X86::GR32RegClass.contains(DestReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3015 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
3016 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
3017 return;
3018 }
3019 }
3020 if (DestReg == X86::EFLAGS) {
3021 if (X86::GR64RegClass.contains(SrcReg)) {
3022 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
3023 .addReg(SrcReg, getKillRegState(KillSrc));
3024 BuildMI(MBB, MI, DL, get(X86::POPF64));
3025 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]3026 }
3027 if (X86::GR32RegClass.contains(SrcReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]3028 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
3029 .addReg(SrcReg, getKillRegState(KillSrc));
3030 BuildMI(MBB, MI, DL, get(X86::POPF32));
3031 return;
3032 }
3033 }
3034
3035 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
3036 << " to " << RI.getName(DestReg) << '\n');
3037 llvm_unreachable("Cannot emit physreg copy instruction");
3038}
3039
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3040static unsigned getLoadStoreRegOpcode(unsigned Reg,
3041 const TargetRegisterClass *RC,
3042 bool isStackAligned,
3043 const TargetMachine &TM,
3044 bool load) {
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3045 if (TM.getSubtarget<X86Subtarget>().hasAVX512()) {
3046 if (X86::VK8RegClass.hasSubClassEq(RC) ||
3047 X86::VK16RegClass.hasSubClassEq(RC))
3048 return load ? X86::KMOVWkm : X86::KMOVWmk;
3049
3050 if (X86::FR32XRegClass.hasSubClassEq(RC))
3051 return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
3052 if (X86::FR64XRegClass.hasSubClassEq(RC))
3053 return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
3054 if (X86::VR128XRegClass.hasSubClassEq(RC) ||
3055 X86::VR256XRegClass.hasSubClassEq(RC) ||
3056 X86::VR512RegClass.hasSubClassEq(RC))
3057 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3058 }
3059
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3060 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3061 switch (RC->getSize()) {
Rafael Espindola6635f982010-07-12 03:43:04 +0000[diff] [blame]3062 default:
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3063 llvm_unreachable("Unknown spill size");
3064 case 1:
3065 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3066 if (TM.getSubtarget<X86Subtarget>().is64Bit())
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3067 // Copying to or from a physical H register on x86-64 requires a NOREX
3068 // move. Otherwise use a normal move.
3069 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3070 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3071 return load ? X86::MOV8rm : X86::MOV8mr;
3072 case 2:
3073 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3074 return load ? X86::MOV16rm : X86::MOV16mr;
3075 case 4:
3076 if (X86::GR32RegClass.hasSubClassEq(RC))
3077 return load ? X86::MOV32rm : X86::MOV32mr;
3078 if (X86::FR32RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3079 return load ?
3080 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3081 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3082 if (X86::RFP32RegClass.hasSubClassEq(RC))
3083 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3084 llvm_unreachable("Unknown 4-byte regclass");
3085 case 8:
3086 if (X86::GR64RegClass.hasSubClassEq(RC))
3087 return load ? X86::MOV64rm : X86::MOV64mr;
3088 if (X86::FR64RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3089 return load ?
3090 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3091 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3092 if (X86::VR64RegClass.hasSubClassEq(RC))
3093 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3094 if (X86::RFP64RegClass.hasSubClassEq(RC))
3095 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3096 llvm_unreachable("Unknown 8-byte regclass");
3097 case 10:
3098 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3099 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]3100 case 16: {
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]3101 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3102 // If stack is realigned we can use aligned stores.
3103 if (isStackAligned)
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]3104 return load ?
3105 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
3106 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3107 else
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]3108 return load ?
3109 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
3110 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
3111 }
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]3112 case 32:
3113 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
3114 // If stack is realigned we can use aligned stores.
3115 if (isStackAligned)
3116 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
3117 else
3118 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3119 case 64:
3120 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3121 if (isStackAligned)
3122 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3123 else
3124 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3125 }
3126}
3127
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3128static unsigned getStoreRegOpcode(unsigned SrcReg,
3129 const TargetRegisterClass *RC,
3130 bool isStackAligned,
3131 TargetMachine &TM) {
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3132 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
3133}
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3134
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]3135
3136static unsigned getLoadRegOpcode(unsigned DestReg,
3137 const TargetRegisterClass *RC,
3138 bool isStackAligned,
3139 const TargetMachine &TM) {
3140 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3141}
3142
3143void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3144 MachineBasicBlock::iterator MI,
3145 unsigned SrcReg, bool isKill, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]3146 const TargetRegisterClass *RC,
3147 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3148 const MachineFunction &MF = *MBB.getParent();
Jakob Stoklund Olesenc3c05ed2010-07-27 04:16:58 +0000[diff] [blame]3149 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
3150 "Stack slot too small for store");
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3151 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3152 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]3153 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3154 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]3155 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3156 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]3157 .addReg(SrcReg, getKillRegState(isKill));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3158}
3159
3160void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
3161 bool isKill,
3162 SmallVectorImpl<MachineOperand> &Addr,
3163 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3164 MachineInstr::mmo_iterator MMOBegin,
3165 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3166 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3167 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3168 bool isAligned = MMOBegin != MMOEnd &&
3169 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3170 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]3171 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]3172 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3173 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3174 MIB.addOperand(Addr[i]);
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]3175 MIB.addReg(SrcReg, getKillRegState(isKill));
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3176 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3177 NewMIs.push_back(MIB);
3178}
3179
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3180
3181void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3182 MachineBasicBlock::iterator MI,
3183 unsigned DestReg, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]3184 const TargetRegisterClass *RC,
3185 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3186 const MachineFunction &MF = *MBB.getParent();
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3187 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3188 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]3189 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3190 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]3191 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3192 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3193}
3194
3195void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]3196 SmallVectorImpl<MachineOperand> &Addr,
3197 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3198 MachineInstr::mmo_iterator MMOBegin,
3199 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3200 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Elena Demikhovsky3ce8dbb2013-08-18 13:08:57 +0000[diff] [blame]3201 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3202 bool isAligned = MMOBegin != MMOEnd &&
3203 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3204 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]3205 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]3206 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3207 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3208 MIB.addOperand(Addr[i]);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3209 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3210 NewMIs.push_back(MIB);
3211}
3212
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3213bool X86InstrInfo::
3214analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3215 int &CmpMask, int &CmpValue) const {
3216 switch (MI->getOpcode()) {
3217 default: break;
3218 case X86::CMP64ri32:
3219 case X86::CMP64ri8:
3220 case X86::CMP32ri:
3221 case X86::CMP32ri8:
3222 case X86::CMP16ri:
3223 case X86::CMP16ri8:
3224 case X86::CMP8ri:
3225 SrcReg = MI->getOperand(0).getReg();
3226 SrcReg2 = 0;
3227 CmpMask = ~0;
3228 CmpValue = MI->getOperand(1).getImm();
3229 return true;
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3230 // A SUB can be used to perform comparison.
3231 case X86::SUB64rm:
3232 case X86::SUB32rm:
3233 case X86::SUB16rm:
3234 case X86::SUB8rm:
3235 SrcReg = MI->getOperand(1).getReg();
3236 SrcReg2 = 0;
3237 CmpMask = ~0;
3238 CmpValue = 0;
3239 return true;
3240 case X86::SUB64rr:
3241 case X86::SUB32rr:
3242 case X86::SUB16rr:
3243 case X86::SUB8rr:
3244 SrcReg = MI->getOperand(1).getReg();
3245 SrcReg2 = MI->getOperand(2).getReg();
3246 CmpMask = ~0;
3247 CmpValue = 0;
3248 return true;
3249 case X86::SUB64ri32:
3250 case X86::SUB64ri8:
3251 case X86::SUB32ri:
3252 case X86::SUB32ri8:
3253 case X86::SUB16ri:
3254 case X86::SUB16ri8:
3255 case X86::SUB8ri:
3256 SrcReg = MI->getOperand(1).getReg();
3257 SrcReg2 = 0;
3258 CmpMask = ~0;
3259 CmpValue = MI->getOperand(2).getImm();
3260 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3261 case X86::CMP64rr:
3262 case X86::CMP32rr:
3263 case X86::CMP16rr:
3264 case X86::CMP8rr:
3265 SrcReg = MI->getOperand(0).getReg();
3266 SrcReg2 = MI->getOperand(1).getReg();
3267 CmpMask = ~0;
3268 CmpValue = 0;
3269 return true;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3270 case X86::TEST8rr:
3271 case X86::TEST16rr:
3272 case X86::TEST32rr:
3273 case X86::TEST64rr:
3274 SrcReg = MI->getOperand(0).getReg();
3275 if (MI->getOperand(1).getReg() != SrcReg) return false;
3276 // Compare against zero.
3277 SrcReg2 = 0;
3278 CmpMask = ~0;
3279 CmpValue = 0;
3280 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3281 }
3282 return false;
3283}
3284
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3285/// isRedundantFlagInstr - check whether the first instruction, whose only
3286/// purpose is to update flags, can be made redundant.
3287/// CMPrr can be made redundant by SUBrr if the operands are the same.
3288/// This function can be extended later on.
3289/// SrcReg, SrcRegs: register operands for FlagI.
3290/// ImmValue: immediate for FlagI if it takes an immediate.
3291inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3292 unsigned SrcReg2, int ImmValue,
3293 MachineInstr *OI) {
3294 if (((FlagI->getOpcode() == X86::CMP64rr &&
3295 OI->getOpcode() == X86::SUB64rr) ||
3296 (FlagI->getOpcode() == X86::CMP32rr &&
3297 OI->getOpcode() == X86::SUB32rr)||
3298 (FlagI->getOpcode() == X86::CMP16rr &&
3299 OI->getOpcode() == X86::SUB16rr)||
3300 (FlagI->getOpcode() == X86::CMP8rr &&
3301 OI->getOpcode() == X86::SUB8rr)) &&
3302 ((OI->getOperand(1).getReg() == SrcReg &&
3303 OI->getOperand(2).getReg() == SrcReg2) ||
3304 (OI->getOperand(1).getReg() == SrcReg2 &&
3305 OI->getOperand(2).getReg() == SrcReg)))
3306 return true;
3307
3308 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3309 OI->getOpcode() == X86::SUB64ri32) ||
3310 (FlagI->getOpcode() == X86::CMP64ri8 &&
3311 OI->getOpcode() == X86::SUB64ri8) ||
3312 (FlagI->getOpcode() == X86::CMP32ri &&
3313 OI->getOpcode() == X86::SUB32ri) ||
3314 (FlagI->getOpcode() == X86::CMP32ri8 &&
3315 OI->getOpcode() == X86::SUB32ri8) ||
3316 (FlagI->getOpcode() == X86::CMP16ri &&
3317 OI->getOpcode() == X86::SUB16ri) ||
3318 (FlagI->getOpcode() == X86::CMP16ri8 &&
3319 OI->getOpcode() == X86::SUB16ri8) ||
3320 (FlagI->getOpcode() == X86::CMP8ri &&
3321 OI->getOpcode() == X86::SUB8ri)) &&
3322 OI->getOperand(1).getReg() == SrcReg &&
3323 OI->getOperand(2).getImm() == ImmValue)
3324 return true;
3325 return false;
3326}
3327
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3328/// isDefConvertible - check whether the definition can be converted
3329/// to remove a comparison against zero.
3330inline static bool isDefConvertible(MachineInstr *MI) {
3331 switch (MI->getOpcode()) {
3332 default: return false;
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]3333
3334 // The shift instructions only modify ZF if their shift count is non-zero.
3335 // N.B.: The processor truncates the shift count depending on the encoding.
3336 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3337 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3338 return getTruncatedShiftCount(MI, 2) != 0;
3339
3340 // Some left shift instructions can be turned into LEA instructions but only
3341 // if their flags aren't used. Avoid transforming such instructions.
3342 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3343 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3344 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3345 return ShAmt != 0;
3346 }
3347
3348 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3349 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3350 return getTruncatedShiftCount(MI, 3) != 0;
3351
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3352 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3353 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3354 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3355 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3356 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
Craig Topper5b08cf72012-12-17 04:55:07 +0000[diff] [blame]3357 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3358 case X86::DEC64_32r: case X86::DEC64_16r:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3359 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3360 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3361 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3362 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3363 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
Craig Topper5b08cf72012-12-17 04:55:07 +0000[diff] [blame]3364 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3365 case X86::INC64_32r: case X86::INC64_16r:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3366 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3367 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3368 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3369 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3370 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3371 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3372 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3373 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3374 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3375 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3376 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3377 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3378 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3379 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3380 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3381 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3382 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3383 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3384 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3385 case X86::ADC32ri: case X86::ADC32ri8:
3386 case X86::ADC32rr: case X86::ADC64ri32:
3387 case X86::ADC64ri8: case X86::ADC64rr:
3388 case X86::SBB32ri: case X86::SBB32ri8:
3389 case X86::SBB32rr: case X86::SBB64ri32:
3390 case X86::SBB64ri8: case X86::SBB64rr:
Craig Topperf3ff6ae2012-12-17 05:12:30 +0000[diff] [blame]3391 case X86::ANDN32rr: case X86::ANDN32rm:
3392 case X86::ANDN64rr: case X86::ANDN64rm:
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3393 case X86::BEXTR32rr: case X86::BEXTR64rr:
3394 case X86::BEXTR32rm: case X86::BEXTR64rm:
3395 case X86::BLSI32rr: case X86::BLSI32rm:
3396 case X86::BLSI64rr: case X86::BLSI64rm:
3397 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3398 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3399 case X86::BLSR32rr: case X86::BLSR32rm:
3400 case X86::BLSR64rr: case X86::BLSR64rm:
3401 case X86::BZHI32rr: case X86::BZHI32rm:
3402 case X86::BZHI64rr: case X86::BZHI64rm:
3403 case X86::LZCNT16rr: case X86::LZCNT16rm:
3404 case X86::LZCNT32rr: case X86::LZCNT32rm:
3405 case X86::LZCNT64rr: case X86::LZCNT64rm:
3406 case X86::POPCNT16rr:case X86::POPCNT16rm:
3407 case X86::POPCNT32rr:case X86::POPCNT32rm:
3408 case X86::POPCNT64rr:case X86::POPCNT64rm:
3409 case X86::TZCNT16rr: case X86::TZCNT16rm:
3410 case X86::TZCNT32rr: case X86::TZCNT32rm:
3411 case X86::TZCNT64rr: case X86::TZCNT64rm:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3412 return true;
3413 }
3414}
3415
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3416/// optimizeCompareInstr - Check if there exists an earlier instruction that
3417/// operates on the same source operands and sets flags in the same way as
3418/// Compare; remove Compare if possible.
3419bool X86InstrInfo::
3420optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3421 int CmpMask, int CmpValue,
3422 const MachineRegisterInfo *MRI) const {
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3423 // Check whether we can replace SUB with CMP.
3424 unsigned NewOpcode = 0;
3425 switch (CmpInstr->getOpcode()) {
3426 default: break;
3427 case X86::SUB64ri32:
3428 case X86::SUB64ri8:
3429 case X86::SUB32ri:
3430 case X86::SUB32ri8:
3431 case X86::SUB16ri:
3432 case X86::SUB16ri8:
3433 case X86::SUB8ri:
3434 case X86::SUB64rm:
3435 case X86::SUB32rm:
3436 case X86::SUB16rm:
3437 case X86::SUB8rm:
3438 case X86::SUB64rr:
3439 case X86::SUB32rr:
3440 case X86::SUB16rr:
3441 case X86::SUB8rr: {
3442 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3443 return false;
3444 // There is no use of the destination register, we can replace SUB with CMP.
3445 switch (CmpInstr->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]3446 default: llvm_unreachable("Unreachable!");
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3447 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3448 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3449 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3450 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3451 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3452 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3453 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3454 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3455 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3456 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3457 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3458 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3459 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3460 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3461 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3462 }
3463 CmpInstr->setDesc(get(NewOpcode));
3464 CmpInstr->RemoveOperand(0);
3465 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3466 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3467 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3468 return false;
3469 }
3470 }
3471
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3472 // Get the unique definition of SrcReg.
3473 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3474 if (!MI) return false;
3475
3476 // CmpInstr is the first instruction of the BB.
3477 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3478
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3479 // If we are comparing against zero, check whether we can use MI to update
3480 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3481 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3482 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3483 !isDefConvertible(MI)))
3484 return false;
3485
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3486 // We are searching for an earlier instruction that can make CmpInstr
3487 // redundant and that instruction will be saved in Sub.
3488 MachineInstr *Sub = NULL;
3489 const TargetRegisterInfo *TRI = &getRegisterInfo();
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3490
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3491 // We iterate backward, starting from the instruction before CmpInstr and
3492 // stop when reaching the definition of a source register or done with the BB.
3493 // RI points to the instruction before CmpInstr.
3494 // If the definition is in this basic block, RE points to the definition;
3495 // otherwise, RE is the rend of the basic block.
3496 MachineBasicBlock::reverse_iterator
3497 RI = MachineBasicBlock::reverse_iterator(I),
3498 RE = CmpInstr->getParent() == MI->getParent() ?
3499 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3500 CmpInstr->getParent()->rend();
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3501 MachineInstr *Movr0Inst = 0;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3502 for (; RI != RE; ++RI) {
3503 MachineInstr *Instr = &*RI;
3504 // Check whether CmpInstr can be made redundant by the current instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3505 if (!IsCmpZero &&
3506 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3507 Sub = Instr;
3508 break;
3509 }
3510
3511 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3512 Instr->readsRegister(X86::EFLAGS, TRI)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3513 // This instruction modifies or uses EFLAGS.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3514
3515 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3516 // They are safe to move up, if the definition to EFLAGS is dead and
3517 // earlier instructions do not read or write EFLAGS.
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]3518 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3519 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3520 Movr0Inst = Instr;
3521 continue;
3522 }
3523
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3524 // We can't remove CmpInstr.
3525 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3526 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3527 }
3528
3529 // Return false if no candidates exist.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3530 if (!IsCmpZero && !Sub)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3531 return false;
3532
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3533 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3534 Sub->getOperand(2).getReg() == SrcReg);
3535
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3536 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3537 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3538 // If we are done with the basic block, we need to check whether EFLAGS is
3539 // live-out.
3540 bool IsSafe = false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3541 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3542 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3543 for (++I; I != E; ++I) {
3544 const MachineInstr &Instr = *I;
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3545 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3546 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3547 // We should check the usage if this instruction uses and updates EFLAGS.
3548 if (!UseEFLAGS && ModifyEFLAGS) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3549 // It is safe to remove CmpInstr if EFLAGS is updated again.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3550 IsSafe = true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3551 break;
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3552 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3553 if (!UseEFLAGS && !ModifyEFLAGS)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3554 continue;
3555
3556 // EFLAGS is used by this instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3557 X86::CondCode OldCC;
3558 bool OpcIsSET = false;
3559 if (IsCmpZero || IsSwapped) {
3560 // We decode the condition code from opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3561 if (Instr.isBranch())
3562 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3563 else {
3564 OldCC = getCondFromSETOpc(Instr.getOpcode());
3565 if (OldCC != X86::COND_INVALID)
3566 OpcIsSET = true;
3567 else
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]3568 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3569 }
3570 if (OldCC == X86::COND_INVALID) return false;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3571 }
3572 if (IsCmpZero) {
3573 switch (OldCC) {
3574 default: break;
3575 case X86::COND_A: case X86::COND_AE:
3576 case X86::COND_B: case X86::COND_BE:
3577 case X86::COND_G: case X86::COND_GE:
3578 case X86::COND_L: case X86::COND_LE:
3579 case X86::COND_O: case X86::COND_NO:
3580 // CF and OF are used, we can't perform this optimization.
3581 return false;
3582 }
3583 } else if (IsSwapped) {
3584 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3585 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3586 // We swap the condition code and synthesize the new opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3587 X86::CondCode NewCC = getSwappedCondition(OldCC);
3588 if (NewCC == X86::COND_INVALID) return false;
3589
3590 // Synthesize the new opcode.
3591 bool HasMemoryOperand = Instr.hasOneMemOperand();
3592 unsigned NewOpc;
3593 if (Instr.isBranch())
3594 NewOpc = GetCondBranchFromCond(NewCC);
3595 else if(OpcIsSET)
3596 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3597 else {
3598 unsigned DstReg = Instr.getOperand(0).getReg();
3599 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3600 HasMemoryOperand);
3601 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3602
3603 // Push the MachineInstr to OpsToUpdate.
3604 // If it is safe to remove CmpInstr, the condition code of these
3605 // instructions will be modified.
3606 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3607 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3608 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3609 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3610 IsSafe = true;
3611 break;
3612 }
3613 }
3614
3615 // If EFLAGS is not killed nor re-defined, we should check whether it is
3616 // live-out. If it is live-out, do not optimize.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3617 if ((IsCmpZero || IsSwapped) && !IsSafe) {
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3618 MachineBasicBlock *MBB = CmpInstr->getParent();
3619 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3620 SE = MBB->succ_end(); SI != SE; ++SI)
3621 if ((*SI)->isLiveIn(X86::EFLAGS))
3622 return false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3623 }
3624
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3625 // The instruction to be updated is either Sub or MI.
3626 Sub = IsCmpZero ? MI : Sub;
David Majnemer5ba473a2013-05-18 01:02:03 +0000[diff] [blame]3627 // Move Movr0Inst to the appropriate place before Sub.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3628 if (Movr0Inst) {
David Majnemer5ba473a2013-05-18 01:02:03 +0000[diff] [blame]3629 // Look backwards until we find a def that doesn't use the current EFLAGS.
3630 Def = Sub;
3631 MachineBasicBlock::reverse_iterator
3632 InsertI = MachineBasicBlock::reverse_iterator(++Def),
3633 InsertE = Sub->getParent()->rend();
3634 for (; InsertI != InsertE; ++InsertI) {
3635 MachineInstr *Instr = &*InsertI;
3636 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3637 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3638 Sub->getParent()->remove(Movr0Inst);
3639 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3640 Movr0Inst);
3641 break;
3642 }
3643 }
3644 if (InsertI == InsertE)
3645 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3646 }
3647
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3648 // Make sure Sub instruction defines EFLAGS and mark the def live.
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3649 unsigned i = 0, e = Sub->getNumOperands();
3650 for (; i != e; ++i) {
3651 MachineOperand &MO = Sub->getOperand(i);
3652 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3653 MO.setIsDead(false);
3654 break;
3655 }
3656 }
3657 assert(i != e && "Unable to locate a def EFLAGS operand");
3658
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3659 CmpInstr->eraseFromParent();
3660
3661 // Modify the condition code of instructions in OpsToUpdate.
3662 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3663 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3664 return true;
3665}
3666
Manman Ren5759d012012-08-02 00:56:42 +0000[diff] [blame]3667/// optimizeLoadInstr - Try to remove the load by folding it to a register
3668/// operand at the use. We fold the load instructions if load defines a virtual
3669/// register, the virtual register is used once in the same BB, and the
3670/// instructions in-between do not load or store, and have no side effects.
3671MachineInstr* X86InstrInfo::
3672optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3673 unsigned &FoldAsLoadDefReg,
3674 MachineInstr *&DefMI) const {
3675 if (FoldAsLoadDefReg == 0)
3676 return 0;
3677 // To be conservative, if there exists another load, clear the load candidate.
3678 if (MI->mayLoad()) {
3679 FoldAsLoadDefReg = 0;
3680 return 0;
3681 }
3682
3683 // Check whether we can move DefMI here.
3684 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3685 assert(DefMI);
3686 bool SawStore = false;
3687 if (!DefMI->isSafeToMove(this, 0, SawStore))
3688 return 0;
3689
3690 // We try to commute MI if possible.
3691 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3692 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3693 // Collect information about virtual register operands of MI.
3694 unsigned SrcOperandId = 0;
3695 bool FoundSrcOperand = false;
3696 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3697 MachineOperand &MO = MI->getOperand(i);
3698 if (!MO.isReg())
3699 continue;
3700 unsigned Reg = MO.getReg();
3701 if (Reg != FoldAsLoadDefReg)
3702 continue;
3703 // Do not fold if we have a subreg use or a def or multiple uses.
3704 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3705 return 0;
3706
3707 SrcOperandId = i;
3708 FoundSrcOperand = true;
3709 }
3710 if (!FoundSrcOperand) return 0;
3711
3712 // Check whether we can fold the def into SrcOperandId.
3713 SmallVector<unsigned, 8> Ops;
3714 Ops.push_back(SrcOperandId);
3715 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3716 if (FoldMI) {
3717 FoldAsLoadDefReg = 0;
3718 return FoldMI;
3719 }
3720
3721 if (Idx == 1) {
3722 // MI was changed but it didn't help, commute it back!
3723 commuteInstruction(MI, false);
3724 return 0;
3725 }
3726
3727 // Check whether we can commute MI and enable folding.
3728 if (MI->isCommutable()) {
3729 MachineInstr *NewMI = commuteInstruction(MI, false);
3730 // Unable to commute.
3731 if (!NewMI) return 0;
3732 if (NewMI != MI) {
3733 // New instruction. It doesn't need to be kept.
3734 NewMI->eraseFromParent();
3735 return 0;
3736 }
3737 }
3738 }
3739 return 0;
3740}
3741
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3742/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3743/// instruction with two undef reads of the register being defined. This is
3744/// used for mapping:
3745/// %xmm4 = V_SET0
3746/// to:
3747/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3748///
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3749static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3750 const MCInstrDesc &Desc) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3751 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3752 unsigned Reg = MIB->getOperand(0).getReg();
3753 MIB->setDesc(Desc);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3754
3755 // MachineInstr::addOperand() will insert explicit operands before any
3756 // implicit operands.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3757 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3758 // But we don't trust that.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3759 assert(MIB->getOperand(1).getReg() == Reg &&
3760 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3761 return true;
3762}
3763
3764bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3765 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3766 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3767 switch (MI->getOpcode()) {
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3768 case X86::SETB_C8r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3769 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3770 case X86::SETB_C16r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3771 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3772 case X86::SETB_C32r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3773 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3774 case X86::SETB_C64r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3775 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3776 case X86::V_SET0:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3777 case X86::FsFLD0SS:
3778 case X86::FsFLD0SD:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3779 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]3780 case X86::AVX_SET0:
3781 assert(HasAVX && "AVX not supported");
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3782 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
Elena Demikhovskyf8f478b2013-08-25 12:54:30 +0000[diff] [blame]3783 case X86::AVX512_512_SET0:
3784 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3785 case X86::V_SETALLONES:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3786 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3787 case X86::AVX2_SETALLONES:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3788 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
Jakob Stoklund Olesen729abd32011-10-08 18:28:28 +0000[diff] [blame]3789 case X86::TEST8ri_NOREX:
3790 MI->setDesc(get(X86::TEST8ri));
3791 return true;
Elena Demikhovskyf8f478b2013-08-25 12:54:30 +0000[diff] [blame]3792 case X86::KSET0W: return Expand2AddrUndef(MIB, get(X86::KXORWrr));
3793 case X86::KSET1B:
3794 case X86::KSET1W: return Expand2AddrUndef(MIB, get(X86::KXNORWrr));
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3795 }
3796 return false;
3797}
3798
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3799static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3800 const SmallVectorImpl<MachineOperand> &MOs,
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3801 MachineInstr *MI,
3802 const TargetInstrInfo &TII) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3803 // Create the base instruction with the memory operand as the first part.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3804 // Omit the implicit operands, something BuildMI can't do.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3805 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3806 MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3807 MachineInstrBuilder MIB(MF, NewMI);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3808 unsigned NumAddrOps = MOs.size();
3809 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3810 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3811 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3812 addOffset(MIB, 0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3813
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3814 // Loop over the rest of the ri operands, converting them over.
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3815 unsigned NumOps = MI->getDesc().getNumOperands()-2;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3816 for (unsigned i = 0; i != NumOps; ++i) {
3817 MachineOperand &MO = MI->getOperand(i+2);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3818 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3819 }
3820 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3821 MachineOperand &MO = MI->getOperand(i);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3822 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3823 }
3824 return MIB;
3825}
3826
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3827static MachineInstr *FuseInst(MachineFunction &MF,
3828 unsigned Opcode, unsigned OpNo,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3829 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3830 MachineInstr *MI, const TargetInstrInfo &TII) {
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3831 // Omit the implicit operands, something BuildMI can't do.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3832 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3833 MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3834 MachineInstrBuilder MIB(MF, NewMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3835
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3836 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3837 MachineOperand &MO = MI->getOperand(i);
3838 if (i == OpNo) {
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3839 assert(MO.isReg() && "Expected to fold into reg operand!");
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3840 unsigned NumAddrOps = MOs.size();
3841 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3842 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3843 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3844 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3845 } else {
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3846 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3847 }
3848 }
3849 return MIB;
3850}
3851
3852static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3853 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3854 MachineInstr *MI) {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3855 MachineFunction &MF = *MI->getParent()->getParent();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3856 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3857
3858 unsigned NumAddrOps = MOs.size();
3859 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3860 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3861 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3862 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3863 return MIB.addImm(0);
3864}
3865
3866MachineInstr*
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3867X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3868 MachineInstr *MI, unsigned i,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3869 const SmallVectorImpl<MachineOperand> &MOs,
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3870 unsigned Size, unsigned Align) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3871 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
Preston Gurdd6be4bf2013-03-27 23:16:18 +0000[diff] [blame]3872 bool isCallRegIndirect = TM.getSubtarget<X86Subtarget>().callRegIndirect();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3873 bool isTwoAddrFold = false;
Preston Gurdd6be4bf2013-03-27 23:16:18 +0000[diff] [blame]3874
3875 // Atom favors register form of call. So, we do not fold loads into calls
3876 // when X86Subtarget is Atom.
3877 if (isCallRegIndirect &&
3878 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
3879 return NULL;
3880 }
3881
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3882 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3883 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]3884 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3885
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]3886 // FIXME: AsmPrinter doesn't know how to handle
3887 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3888 if (MI->getOpcode() == X86::ADD32ri &&
3889 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3890 return NULL;
3891
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3892 MachineInstr *NewMI = NULL;
3893 // Folding a memory location into the two-address part of a two-address
3894 // instruction is different than folding it other places. It requires
3895 // replacing the *two* registers with the memory location.
3896 if (isTwoAddr && NumOps >= 2 && i < 2 &&
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3897 MI->getOperand(0).isReg() &&
3898 MI->getOperand(1).isReg() &&
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3899 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3900 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3901 isTwoAddrFold = true;
3902 } else if (i == 0) { // If operand 0
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]3903 if (MI->getOpcode() == X86::MOV32r0) {
3904 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
3905 if (NewMI)
3906 return NewMI;
Craig Topperf9115972012-08-23 04:57:36 +0000[diff] [blame]3907 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3908
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3909 OpcodeTablePtr = &RegOp2MemOpTable0;
3910 } else if (i == 1) {
3911 OpcodeTablePtr = &RegOp2MemOpTable1;
3912 } else if (i == 2) {
3913 OpcodeTablePtr = &RegOp2MemOpTable2;
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]3914 } else if (i == 3) {
3915 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3916 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3917
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3918 // If table selected...
3919 if (OpcodeTablePtr) {
3920 // Find the Opcode to fuse
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3921 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3922 OpcodeTablePtr->find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3923 if (I != OpcodeTablePtr->end()) {
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3924 unsigned Opcode = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]3925 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3926 if (Align < MinAlign)
3927 return NULL;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3928 bool NarrowToMOV32rm = false;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3929 if (Size) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]3930 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3931 if (Size < RCSize) {
3932 // Check if it's safe to fold the load. If the size of the object is
3933 // narrower than the load width, then it's not.
3934 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3935 return NULL;
3936 // If this is a 64-bit load, but the spill slot is 32, then we can do
3937 // a 32-bit load which is implicitly zero-extended. This likely is due
3938 // to liveintervalanalysis remat'ing a load from stack slot.
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3939 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3940 return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3941 Opcode = X86::MOV32rm;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3942 NarrowToMOV32rm = true;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3943 }
3944 }
3945
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3946 if (isTwoAddrFold)
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3947 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3948 else
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3949 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3950
3951 if (NarrowToMOV32rm) {
3952 // If this is the special case where we use a MOV32rm to load a 32-bit
3953 // value and zero-extend the top bits. Change the destination register
3954 // to a 32-bit one.
3955 unsigned DstReg = NewMI->getOperand(0).getReg();
3956 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3957 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3958 X86::sub_32bit));
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3959 else
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3960 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3961 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3962 return NewMI;
3963 }
3964 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3965
3966 // No fusion
Jakob Stoklund Olesen51702ec2010-07-09 20:43:09 +0000[diff] [blame]3967 if (PrintFailedFusing && !MI->isCopy())
David Greened589daf2010-01-05 01:29:29 +0000[diff] [blame]3968 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3969 return NULL;
3970}
3971
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3972/// hasPartialRegUpdate - Return true for all instructions that only update
3973/// the first 32 or 64-bits of the destination register and leave the rest
3974/// unmodified. This can be used to avoid folding loads if the instructions
3975/// only update part of the destination register, and the non-updated part is
3976/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3977/// instructions breaks the partial register dependency and it can improve
3978/// performance. e.g.:
3979///
3980/// movss (%rdi), %xmm0
3981/// cvtss2sd %xmm0, %xmm0
3982///
3983/// Instead of
3984/// cvtss2sd (%rdi), %xmm0
3985///
Bruno Cardoso Lopes7b435682011-09-15 23:04:24 +0000[diff] [blame]3986/// FIXME: This should be turned into a TSFlags.
3987///
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3988static bool hasPartialRegUpdate(unsigned Opcode) {
3989 switch (Opcode) {
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]3990 case X86::CVTSI2SSrr:
3991 case X86::CVTSI2SS64rr:
3992 case X86::CVTSI2SDrr:
3993 case X86::CVTSI2SD64rr:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3994 case X86::CVTSD2SSrr:
3995 case X86::Int_CVTSD2SSrr:
3996 case X86::CVTSS2SDrr:
3997 case X86::Int_CVTSS2SDrr:
3998 case X86::RCPSSr:
3999 case X86::RCPSSr_Int:
4000 case X86::ROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]4001 case X86::ROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4002 case X86::ROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]4003 case X86::ROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4004 case X86::RSQRTSSr:
4005 case X86::RSQRTSSr_Int:
4006 case X86::SQRTSSr:
4007 case X86::SQRTSSr_Int:
4008 // AVX encoded versions
4009 case X86::VCVTSD2SSrr:
4010 case X86::Int_VCVTSD2SSrr:
4011 case X86::VCVTSS2SDrr:
4012 case X86::Int_VCVTSS2SDrr:
4013 case X86::VRCPSSr:
4014 case X86::VROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]4015 case X86::VROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4016 case X86::VROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]4017 case X86::VROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4018 case X86::VRSQRTSSr:
4019 case X86::VSQRTSSr:
4020 return true;
4021 }
4022
4023 return false;
4024}
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4025
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]4026/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
4027/// instructions we would like before a partial register update.
4028unsigned X86InstrInfo::
4029getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
4030 const TargetRegisterInfo *TRI) const {
4031 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
4032 return 0;
4033
4034 // If MI is marked as reading Reg, the partial register update is wanted.
4035 const MachineOperand &MO = MI->getOperand(0);
4036 unsigned Reg = MO.getReg();
4037 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4038 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
4039 return 0;
4040 } else {
4041 if (MI->readsRegister(Reg, TRI))
4042 return 0;
4043 }
4044
4045 // If any of the preceding 16 instructions are reading Reg, insert a
4046 // dependency breaking instruction. The magic number is based on a few
4047 // Nehalem experiments.
4048 return 16;
4049}
4050
4051void X86InstrInfo::
4052breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
4053 const TargetRegisterInfo *TRI) const {
4054 unsigned Reg = MI->getOperand(OpNum).getReg();
4055 if (X86::VR128RegClass.contains(Reg)) {
4056 // These instructions are all floating point domain, so xorps is the best
4057 // choice.
4058 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
4059 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
4060 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
4061 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4062 } else if (X86::VR256RegClass.contains(Reg)) {
4063 // Use vxorps to clear the full ymm register.
4064 // It wants to read and write the xmm sub-register.
4065 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4066 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
4067 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
4068 .addReg(Reg, RegState::ImplicitDefine);
4069 } else
4070 return;
4071 MI->addRegisterKilled(Reg, TRI, true);
4072}
4073
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]4074MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4075 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]4076 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]4077 int FrameIndex) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4078 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4079 if (NoFusing) return NULL;
4080
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4081 // Unless optimizing for size, don't fold to avoid partial
4082 // register update stalls
Bill Wendling698e84f2012-12-30 10:32:01 +0000[diff] [blame]4083 if (!MF.getFunction()->getAttributes().
4084 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4085 hasPartialRegUpdate(MI->getOpcode()))
4086 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]4087
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]4088 const MachineFrameInfo *MFI = MF.getFrameInfo();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4089 unsigned Size = MFI->getObjectSize(FrameIndex);
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]4090 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4091 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4092 unsigned NewOpc = 0;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4093 unsigned RCSize = 0;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4094 switch (MI->getOpcode()) {
4095 default: return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4096 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
Dan Gohman887dd1c2010-05-18 21:42:03 +0000[diff] [blame]4097 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
4098 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
4099 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4100 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4101 // Check if it's safe to fold the load. If the size of the object is
4102 // narrower than the load width, then it's not.
4103 if (Size < RCSize)
4104 return NULL;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4105 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4106 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4107 MI->getOperand(1).ChangeToImmediate(0);
4108 } else if (Ops.size() != 1)
4109 return NULL;
4110
4111 SmallVector<MachineOperand,4> MOs;
4112 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4113 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4114}
4115
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]4116MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4117 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]4118 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]4119 MachineInstr *LoadMI) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4120 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4121 if (NoFusing) return NULL;
4122
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4123 // Unless optimizing for size, don't fold to avoid partial
4124 // register update stalls
Bill Wendling698e84f2012-12-30 10:32:01 +0000[diff] [blame]4125 if (!MF.getFunction()->getAttributes().
4126 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4127 hasPartialRegUpdate(MI->getOpcode()))
4128 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]4129
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]4130 // Determine the alignment of the load.
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]4131 unsigned Alignment = 0;
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]4132 if (LoadMI->hasOneMemOperand())
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]4133 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4134 else
4135 switch (LoadMI->getOpcode()) {
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]4136 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4137 case X86::AVX_SET0:
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4138 Alignment = 32;
4139 break;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4140 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4141 case X86::V_SETALLONES:
4142 Alignment = 16;
4143 break;
4144 case X86::FsFLD0SD:
4145 Alignment = 8;
4146 break;
4147 case X86::FsFLD0SS:
4148 Alignment = 4;
4149 break;
4150 default:
Eli Friedman87ef3872011-06-10 01:13:01 +0000[diff] [blame]4151 return 0;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4152 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4153 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4154 unsigned NewOpc = 0;
4155 switch (MI->getOpcode()) {
4156 default: return NULL;
4157 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4158 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
4159 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
4160 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4161 }
4162 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4163 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4164 MI->getOperand(1).ChangeToImmediate(0);
4165 } else if (Ops.size() != 1)
4166 return NULL;
4167
Jakob Stoklund Olesen9c473e42010-08-11 23:08:22 +0000[diff] [blame]4168 // Make sure the subregisters match.
4169 // Otherwise we risk changing the size of the load.
4170 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4171 return NULL;
4172
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4173 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4174 switch (LoadMI->getOpcode()) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4175 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4176 case X86::V_SETALLONES:
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]4177 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4178 case X86::AVX_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4179 case X86::FsFLD0SD:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4180 case X86::FsFLD0SS: {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4181 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4182 // Create a constant-pool entry and operands to load from it.
4183
Dan Gohman772952f2010-03-09 03:01:40 +0000[diff] [blame]4184 // Medium and large mode can't fold loads this way.
4185 if (TM.getCodeModel() != CodeModel::Small &&
4186 TM.getCodeModel() != CodeModel::Kernel)
4187 return NULL;
4188
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4189 // x86-32 PIC requires a PIC base register for constant pools.
4190 unsigned PICBase = 0;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4191 if (TM.getRelocationModel() == Reloc::PIC_) {
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]4192 if (TM.getSubtarget<X86Subtarget>().is64Bit())
4193 PICBase = X86::RIP;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4194 else
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4195 // FIXME: PICBase = getGlobalBaseReg(&MF);
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]4196 // This doesn't work for several reasons.
4197 // 1. GlobalBaseReg may have been spilled.
4198 // 2. It may not be live at MI.
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4199 return NULL;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4200 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4201
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4202 // Create a constant-pool entry.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4203 MachineConstantPool &MCP = *MF.getConstantPool();
Chris Lattner229907c2011-07-18 04:54:35 +0000[diff] [blame]4204 Type *Ty;
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4205 unsigned Opc = LoadMI->getOpcode();
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4206 if (Opc == X86::FsFLD0SS)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4207 Ty = Type::getFloatTy(MF.getFunction()->getContext());
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4208 else if (Opc == X86::FsFLD0SD)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4209 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4210 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
Craig Toppera4c5a472012-01-13 06:12:41 +0000[diff] [blame]4211 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4212 else
4213 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]4214
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]4215 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]4216 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4217 Constant::getNullValue(Ty);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4218 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4219
4220 // Create operands to load from the constant pool entry.
4221 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4222 MOs.push_back(MachineOperand::CreateImm(1));
4223 MOs.push_back(MachineOperand::CreateReg(0, false));
4224 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]4225 MOs.push_back(MachineOperand::CreateReg(0, false));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4226 break;
4227 }
4228 default: {
Manman Ren5b462822012-11-27 18:09:26 +0000[diff] [blame]4229 if ((LoadMI->getOpcode() == X86::MOVSSrm ||
4230 LoadMI->getOpcode() == X86::VMOVSSrm) &&
4231 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4232 > 4)
4233 // These instructions only load 32 bits, we can't fold them if the
4234 // destination register is wider than 32 bits (4 bytes).
4235 return NULL;
4236 if ((LoadMI->getOpcode() == X86::MOVSDrm ||
4237 LoadMI->getOpcode() == X86::VMOVSDrm) &&
4238 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4239 > 8)
4240 // These instructions only load 64 bits, we can't fold them if the
4241 // destination register is wider than 64 bits (8 bytes).
4242 return NULL;
4243
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4244 // Folding a normal load. Just copy the load's address operands.
4245 unsigned NumOps = LoadMI->getDesc().getNumOperands();
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4246 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4247 MOs.push_back(LoadMI->getOperand(i));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4248 break;
4249 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4250 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4251 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4252}
4253
4254
Dan Gohman33332bc2008-10-16 01:49:15 +0000[diff] [blame]4255bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4256 const SmallVectorImpl<unsigned> &Ops) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4257 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4258 if (NoFusing) return 0;
4259
4260 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4261 switch (MI->getOpcode()) {
4262 default: return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4263 case X86::TEST8rr:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4264 case X86::TEST16rr:
4265 case X86::TEST32rr:
4266 case X86::TEST64rr:
4267 return true;
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]4268 case X86::ADD32ri:
4269 // FIXME: AsmPrinter doesn't know how to handle
4270 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4271 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4272 return false;
4273 break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4274 }
4275 }
4276
4277 if (Ops.size() != 1)
4278 return false;
4279
4280 unsigned OpNum = Ops[0];
4281 unsigned Opc = MI->getOpcode();
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]4282 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4283 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4284 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4285
4286 // Folding a memory location into the two-address part of a two-address
4287 // instruction is different than folding it other places. It requires
4288 // replacing the *two* registers with the memory location.
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4289 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4290 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4291 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4292 } else if (OpNum == 0) { // If operand 0
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]4293 if (Opc == X86::MOV32r0)
4294 return true;
4295
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4296 OpcodeTablePtr = &RegOp2MemOpTable0;
4297 } else if (OpNum == 1) {
4298 OpcodeTablePtr = &RegOp2MemOpTable1;
4299 } else if (OpNum == 2) {
4300 OpcodeTablePtr = &RegOp2MemOpTable2;
Craig Topper7573c8f2012-08-31 22:12:16 +0000[diff] [blame]4301 } else if (OpNum == 3) {
4302 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4303 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4304
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]4305 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4306 return true;
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]4307 return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4308}
4309
4310bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4311 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4312 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4313 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4314 MemOp2RegOpTable.find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4315 if (I == MemOp2RegOpTable.end())
4316 return false;
4317 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4318 unsigned Index = I->second.second & TB_INDEX_MASK;
4319 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4320 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4321 if (UnfoldLoad && !FoldedLoad)
4322 return false;
4323 UnfoldLoad &= FoldedLoad;
4324 if (UnfoldStore && !FoldedStore)
4325 return false;
4326 UnfoldStore &= FoldedStore;
4327
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4328 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4329 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4330 if (!MI->hasOneMemOperand() &&
4331 RC == &X86::VR128RegClass &&
4332 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4333 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4334 // conservatively assume the address is unaligned. That's bad for
4335 // performance.
4336 return false;
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4337 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4338 SmallVector<MachineOperand,2> BeforeOps;
4339 SmallVector<MachineOperand,2> AfterOps;
4340 SmallVector<MachineOperand,4> ImpOps;
4341 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4342 MachineOperand &Op = MI->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4343 if (i >= Index && i < Index + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4344 AddrOps.push_back(Op);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4345 else if (Op.isReg() && Op.isImplicit())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4346 ImpOps.push_back(Op);
4347 else if (i < Index)
4348 BeforeOps.push_back(Op);
4349 else if (i > Index)
4350 AfterOps.push_back(Op);
4351 }
4352
4353 // Emit the load instruction.
4354 if (UnfoldLoad) {
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4355 std::pair<MachineInstr::mmo_iterator,
4356 MachineInstr::mmo_iterator> MMOs =
4357 MF.extractLoadMemRefs(MI->memoperands_begin(),
4358 MI->memoperands_end());
4359 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4360 if (UnfoldStore) {
4361 // Address operands cannot be marked isKill.
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4362 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4363 MachineOperand &MO = NewMIs[0]->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4364 if (MO.isReg())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4365 MO.setIsKill(false);
4366 }
4367 }
4368 }
4369
4370 // Emit the data processing instruction.
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4371 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]4372 MachineInstrBuilder MIB(MF, DataMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4373
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4374 if (FoldedStore)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4375 MIB.addReg(Reg, RegState::Define);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4376 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4377 MIB.addOperand(BeforeOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4378 if (FoldedLoad)
4379 MIB.addReg(Reg);
4380 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4381 MIB.addOperand(AfterOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4382 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4383 MachineOperand &MO = ImpOps[i];
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4384 MIB.addReg(MO.getReg(),
4385 getDefRegState(MO.isDef()) |
4386 RegState::Implicit |
4387 getKillRegState(MO.isKill()) |
Evan Cheng0dc101b2009-06-30 08:49:04 +0000[diff] [blame]4388 getDeadRegState(MO.isDead()) |
4389 getUndefRegState(MO.isUndef()));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4390 }
4391 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4392 switch (DataMI->getOpcode()) {
4393 default: break;
4394 case X86::CMP64ri32:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4395 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4396 case X86::CMP32ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4397 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4398 case X86::CMP16ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4399 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4400 case X86::CMP8ri: {
4401 MachineOperand &MO0 = DataMI->getOperand(0);
4402 MachineOperand &MO1 = DataMI->getOperand(1);
4403 if (MO1.getImm() == 0) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4404 unsigned NewOpc;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4405 switch (DataMI->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4406 default: llvm_unreachable("Unreachable!");
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4407 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4408 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4409 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4410 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4411 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4412 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4413 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4414 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4415 DataMI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4416 MO1.ChangeToRegister(MO0.getReg(), false);
4417 }
4418 }
4419 }
4420 NewMIs.push_back(DataMI);
4421
4422 // Emit the store instruction.
4423 if (UnfoldStore) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4424 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4425 std::pair<MachineInstr::mmo_iterator,
4426 MachineInstr::mmo_iterator> MMOs =
4427 MF.extractStoreMemRefs(MI->memoperands_begin(),
4428 MI->memoperands_end());
4429 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4430 }
4431
4432 return true;
4433}
4434
4435bool
4436X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4437 SmallVectorImpl<SDNode*> &NewNodes) const {
Dan Gohman17059682008-07-17 19:10:17 +0000[diff] [blame]4438 if (!N->isMachineOpcode())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4439 return false;
4440
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4441 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4442 MemOp2RegOpTable.find(N->getMachineOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4443 if (I == MemOp2RegOpTable.end())
4444 return false;
4445 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4446 unsigned Index = I->second.second & TB_INDEX_MASK;
4447 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4448 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4449 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4450 MachineFunction &MF = DAG.getMachineFunction();
4451 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4452 unsigned NumDefs = MCID.NumDefs;
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4453 std::vector<SDValue> AddrOps;
4454 std::vector<SDValue> BeforeOps;
4455 std::vector<SDValue> AfterOps;
Andrew Trickef9de2a2013-05-25 02:42:55 +0000[diff] [blame]4456 SDLoc dl(N);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4457 unsigned NumOps = N->getNumOperands();
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]4458 for (unsigned i = 0; i != NumOps-1; ++i) {
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4459 SDValue Op = N->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4460 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4461 AddrOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4462 else if (i < Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4463 BeforeOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4464 else if (i > Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4465 AfterOps.push_back(Op);
4466 }
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4467 SDValue Chain = N->getOperand(NumOps-1);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4468 AddrOps.push_back(Chain);
4469
4470 // Emit the load instruction.
4471 SDNode *Load = 0;
4472 if (FoldedLoad) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4473 EVT VT = *RC->vt_begin();
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4474 std::pair<MachineInstr::mmo_iterator,
4475 MachineInstr::mmo_iterator> MMOs =
4476 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4477 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4478 if (!(*MMOs.first) &&
4479 RC == &X86::VR128RegClass &&
4480 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4481 // Do not introduce a slow unaligned load.
4482 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4483 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4484 bool isAligned = (*MMOs.first) &&
4485 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4486 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4487 VT, MVT::Other, AddrOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4488 NewNodes.push_back(Load);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4489
4490 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4491 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4492 }
4493
4494 // Emit the data processing instruction.
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4495 std::vector<EVT> VTs;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4496 const TargetRegisterClass *DstRC = 0;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4497 if (MCID.getNumDefs() > 0) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4498 DstRC = getRegClass(MCID, 0, &RI, MF);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4499 VTs.push_back(*DstRC->vt_begin());
4500 }
4501 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4502 EVT VT = N->getValueType(i);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4503 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4504 VTs.push_back(VT);
4505 }
4506 if (Load)
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4507 BeforeOps.push_back(SDValue(Load, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4508 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4509 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4510 NewNodes.push_back(NewNode);
4511
4512 // Emit the store instruction.
4513 if (FoldedStore) {
4514 AddrOps.pop_back();
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4515 AddrOps.push_back(SDValue(NewNode, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4516 AddrOps.push_back(Chain);
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4517 std::pair<MachineInstr::mmo_iterator,
4518 MachineInstr::mmo_iterator> MMOs =
4519 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4520 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4521 if (!(*MMOs.first) &&
4522 RC == &X86::VR128RegClass &&
4523 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4524 // Do not introduce a slow unaligned store.
4525 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4526 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4527 bool isAligned = (*MMOs.first) &&
4528 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4529 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4530 isAligned, TM),
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4531 dl, MVT::Other, AddrOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4532 NewNodes.push_back(Store);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4533
4534 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4535 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4536 }
4537
4538 return true;
4539}
4540
4541unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4542 bool UnfoldLoad, bool UnfoldStore,
4543 unsigned *LoadRegIndex) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4544 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4545 MemOp2RegOpTable.find(Opc);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4546 if (I == MemOp2RegOpTable.end())
4547 return 0;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4548 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4549 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4550 if (UnfoldLoad && !FoldedLoad)
4551 return 0;
4552 if (UnfoldStore && !FoldedStore)
4553 return 0;
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4554 if (LoadRegIndex)
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4555 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4556 return I->second.first;
4557}
4558
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4559bool
4560X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4561 int64_t &Offset1, int64_t &Offset2) const {
4562 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4563 return false;
4564 unsigned Opc1 = Load1->getMachineOpcode();
4565 unsigned Opc2 = Load2->getMachineOpcode();
4566 switch (Opc1) {
4567 default: return false;
4568 case X86::MOV8rm:
4569 case X86::MOV16rm:
4570 case X86::MOV32rm:
4571 case X86::MOV64rm:
4572 case X86::LD_Fp32m:
4573 case X86::LD_Fp64m:
4574 case X86::LD_Fp80m:
4575 case X86::MOVSSrm:
4576 case X86::MOVSDrm:
4577 case X86::MMX_MOVD64rm:
4578 case X86::MMX_MOVQ64rm:
4579 case X86::FsMOVAPSrm:
4580 case X86::FsMOVAPDrm:
4581 case X86::MOVAPSrm:
4582 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4583 case X86::MOVAPDrm:
4584 case X86::MOVDQArm:
4585 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4586 // AVX load instructions
4587 case X86::VMOVSSrm:
4588 case X86::VMOVSDrm:
4589 case X86::FsVMOVAPSrm:
4590 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4591 case X86::VMOVAPSrm:
4592 case X86::VMOVUPSrm:
4593 case X86::VMOVAPDrm:
4594 case X86::VMOVDQArm:
4595 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4596 case X86::VMOVAPSYrm:
4597 case X86::VMOVUPSYrm:
4598 case X86::VMOVAPDYrm:
4599 case X86::VMOVDQAYrm:
4600 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4601 break;
4602 }
4603 switch (Opc2) {
4604 default: return false;
4605 case X86::MOV8rm:
4606 case X86::MOV16rm:
4607 case X86::MOV32rm:
4608 case X86::MOV64rm:
4609 case X86::LD_Fp32m:
4610 case X86::LD_Fp64m:
4611 case X86::LD_Fp80m:
4612 case X86::MOVSSrm:
4613 case X86::MOVSDrm:
4614 case X86::MMX_MOVD64rm:
4615 case X86::MMX_MOVQ64rm:
4616 case X86::FsMOVAPSrm:
4617 case X86::FsMOVAPDrm:
4618 case X86::MOVAPSrm:
4619 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4620 case X86::MOVAPDrm:
4621 case X86::MOVDQArm:
4622 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4623 // AVX load instructions
4624 case X86::VMOVSSrm:
4625 case X86::VMOVSDrm:
4626 case X86::FsVMOVAPSrm:
4627 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4628 case X86::VMOVAPSrm:
4629 case X86::VMOVUPSrm:
4630 case X86::VMOVAPDrm:
4631 case X86::VMOVDQArm:
4632 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4633 case X86::VMOVAPSYrm:
4634 case X86::VMOVUPSYrm:
4635 case X86::VMOVAPDYrm:
4636 case X86::VMOVDQAYrm:
4637 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4638 break;
4639 }
4640
4641 // Check if chain operands and base addresses match.
4642 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4643 Load1->getOperand(5) != Load2->getOperand(5))
4644 return false;
4645 // Segment operands should match as well.
4646 if (Load1->getOperand(4) != Load2->getOperand(4))
4647 return false;
4648 // Scale should be 1, Index should be Reg0.
4649 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4650 Load1->getOperand(2) == Load2->getOperand(2)) {
4651 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4652 return false;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4653
4654 // Now let's examine the displacements.
4655 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4656 isa<ConstantSDNode>(Load2->getOperand(3))) {
4657 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4658 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4659 return true;
4660 }
4661 }
4662 return false;
4663}
4664
4665bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4666 int64_t Offset1, int64_t Offset2,
4667 unsigned NumLoads) const {
4668 assert(Offset2 > Offset1);
4669 if ((Offset2 - Offset1) / 8 > 64)
4670 return false;
4671
4672 unsigned Opc1 = Load1->getMachineOpcode();
4673 unsigned Opc2 = Load2->getMachineOpcode();
4674 if (Opc1 != Opc2)
4675 return false; // FIXME: overly conservative?
4676
4677 switch (Opc1) {
4678 default: break;
4679 case X86::LD_Fp32m:
4680 case X86::LD_Fp64m:
4681 case X86::LD_Fp80m:
4682 case X86::MMX_MOVD64rm:
4683 case X86::MMX_MOVQ64rm:
4684 return false;
4685 }
4686
4687 EVT VT = Load1->getValueType(0);
4688 switch (VT.getSimpleVT().SimpleTy) {
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4689 default:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4690 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4691 // have 16 of them to play with.
4692 if (TM.getSubtargetImpl()->is64Bit()) {
4693 if (NumLoads >= 3)
4694 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4695 } else if (NumLoads) {
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4696 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4697 }
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4698 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4699 case MVT::i8:
4700 case MVT::i16:
4701 case MVT::i32:
4702 case MVT::i64:
Evan Cheng16cf9342010-01-22 23:49:11 +0000[diff] [blame]4703 case MVT::f32:
4704 case MVT::f64:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4705 if (NumLoads)
4706 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4707 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4708 }
4709
4710 return true;
4711}
4712
Andrew Trick47740de2013-06-23 09:00:28 +0000[diff] [blame]4713bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First,
4714 MachineInstr *Second) const {
4715 // Check if this processor supports macro-fusion. Since this is a minor
4716 // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
4717 // proxy for SandyBridge+.
4718 if (!TM.getSubtarget<X86Subtarget>().hasAVX())
4719 return false;
4720
4721 enum {
4722 FuseTest,
4723 FuseCmp,
4724 FuseInc
4725 } FuseKind;
4726
4727 switch(Second->getOpcode()) {
4728 default:
4729 return false;
4730 case X86::JE_4:
4731 case X86::JNE_4:
4732 case X86::JL_4:
4733 case X86::JLE_4:
4734 case X86::JG_4:
4735 case X86::JGE_4:
4736 FuseKind = FuseInc;
4737 break;
4738 case X86::JB_4:
4739 case X86::JBE_4:
4740 case X86::JA_4:
4741 case X86::JAE_4:
4742 FuseKind = FuseCmp;
4743 break;
4744 case X86::JS_4:
4745 case X86::JNS_4:
4746 case X86::JP_4:
4747 case X86::JNP_4:
4748 case X86::JO_4:
4749 case X86::JNO_4:
4750 FuseKind = FuseTest;
4751 break;
4752 }
4753 switch (First->getOpcode()) {
4754 default:
4755 return false;
4756 case X86::TEST8rr:
4757 case X86::TEST16rr:
4758 case X86::TEST32rr:
4759 case X86::TEST64rr:
4760 case X86::TEST8ri:
4761 case X86::TEST16ri:
4762 case X86::TEST32ri:
4763 case X86::TEST32i32:
4764 case X86::TEST64i32:
4765 case X86::TEST64ri32:
4766 case X86::TEST8rm:
4767 case X86::TEST16rm:
4768 case X86::TEST32rm:
4769 case X86::TEST64rm:
4770 case X86::AND16i16:
4771 case X86::AND16ri:
4772 case X86::AND16ri8:
4773 case X86::AND16rm:
4774 case X86::AND16rr:
4775 case X86::AND32i32:
4776 case X86::AND32ri:
4777 case X86::AND32ri8:
4778 case X86::AND32rm:
4779 case X86::AND32rr:
4780 case X86::AND64i32:
4781 case X86::AND64ri32:
4782 case X86::AND64ri8:
4783 case X86::AND64rm:
4784 case X86::AND64rr:
4785 case X86::AND8i8:
4786 case X86::AND8ri:
4787 case X86::AND8rm:
4788 case X86::AND8rr:
4789 return true;
4790 case X86::CMP16i16:
4791 case X86::CMP16ri:
4792 case X86::CMP16ri8:
4793 case X86::CMP16rm:
4794 case X86::CMP16rr:
4795 case X86::CMP32i32:
4796 case X86::CMP32ri:
4797 case X86::CMP32ri8:
4798 case X86::CMP32rm:
4799 case X86::CMP32rr:
4800 case X86::CMP64i32:
4801 case X86::CMP64ri32:
4802 case X86::CMP64ri8:
4803 case X86::CMP64rm:
4804 case X86::CMP64rr:
4805 case X86::CMP8i8:
4806 case X86::CMP8ri:
4807 case X86::CMP8rm:
4808 case X86::CMP8rr:
4809 case X86::ADD16i16:
4810 case X86::ADD16ri:
4811 case X86::ADD16ri8:
4812 case X86::ADD16ri8_DB:
4813 case X86::ADD16ri_DB:
4814 case X86::ADD16rm:
4815 case X86::ADD16rr:
4816 case X86::ADD16rr_DB:
4817 case X86::ADD32i32:
4818 case X86::ADD32ri:
4819 case X86::ADD32ri8:
4820 case X86::ADD32ri8_DB:
4821 case X86::ADD32ri_DB:
4822 case X86::ADD32rm:
4823 case X86::ADD32rr:
4824 case X86::ADD32rr_DB:
4825 case X86::ADD64i32:
4826 case X86::ADD64ri32:
4827 case X86::ADD64ri32_DB:
4828 case X86::ADD64ri8:
4829 case X86::ADD64ri8_DB:
4830 case X86::ADD64rm:
4831 case X86::ADD64rr:
4832 case X86::ADD64rr_DB:
4833 case X86::ADD8i8:
4834 case X86::ADD8mi:
4835 case X86::ADD8mr:
4836 case X86::ADD8ri:
4837 case X86::ADD8rm:
4838 case X86::ADD8rr:
4839 case X86::SUB16i16:
4840 case X86::SUB16ri:
4841 case X86::SUB16ri8:
4842 case X86::SUB16rm:
4843 case X86::SUB16rr:
4844 case X86::SUB32i32:
4845 case X86::SUB32ri:
4846 case X86::SUB32ri8:
4847 case X86::SUB32rm:
4848 case X86::SUB32rr:
4849 case X86::SUB64i32:
4850 case X86::SUB64ri32:
4851 case X86::SUB64ri8:
4852 case X86::SUB64rm:
4853 case X86::SUB64rr:
4854 case X86::SUB8i8:
4855 case X86::SUB8ri:
4856 case X86::SUB8rm:
4857 case X86::SUB8rr:
4858 return FuseKind == FuseCmp || FuseKind == FuseInc;
4859 case X86::INC16r:
4860 case X86::INC32r:
4861 case X86::INC64_16r:
4862 case X86::INC64_32r:
4863 case X86::INC64r:
4864 case X86::INC8r:
4865 case X86::DEC16r:
4866 case X86::DEC32r:
4867 case X86::DEC64_16r:
4868 case X86::DEC64_32r:
4869 case X86::DEC64r:
4870 case X86::DEC8r:
4871 return FuseKind == FuseInc;
4872 }
4873}
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4874
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4875bool X86InstrInfo::
Owen Anderson4f6bf042008-08-14 22:49:33 +0000[diff] [blame]4876ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4877 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4878 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]4879 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4880 return true;
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4881 Cond[0].setImm(GetOppositeBranchCondition(CC));
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4882 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4883}
4884
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4885bool X86InstrInfo::
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4886isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4887 // FIXME: Return false for x87 stack register classes for now. We can't
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4888 // allow any loads of these registers before FpGet_ST0_80.
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4889 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4890 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4891}
4892
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4893/// getGlobalBaseReg - Return a virtual register initialized with the
4894/// the global base register value. Output instructions required to
4895/// initialize the register in the function entry block, if necessary.
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4896///
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4897/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4898///
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4899unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4900 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4901 "X86-64 PIC uses RIP relative addressing");
4902
4903 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4904 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4905 if (GlobalBaseReg != 0)
4906 return GlobalBaseReg;
4907
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4908 // Create the register. The code to initialize it is inserted
4909 // later, by the CGBR pass (below).
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4910 MachineRegisterInfo &RegInfo = MF->getRegInfo();
Jakob Stoklund Olesen38dcd592012-05-20 18:43:00 +0000[diff] [blame]4911 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4912 X86FI->setGlobalBaseReg(GlobalBaseReg);
4913 return GlobalBaseReg;
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4914}
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4915
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4916// These are the replaceable SSE instructions. Some of these have Int variants
4917// that we don't include here. We don't want to replace instructions selected
4918// by intrinsics.
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4919static const uint16_t ReplaceableInstrs[][3] = {
Bruno Cardoso Lopes1401e042010-08-12 02:08:52 +0000[diff] [blame]4920 //PackedSingle PackedDouble PackedInt
Jakob Stoklund Olesendbff4e82010-03-30 22:46:53 +0000[diff] [blame]4921 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4922 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4923 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4924 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4925 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4926 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4927 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4928 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4929 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4930 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4931 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4932 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4933 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4934 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4935 // AVX 128-bit support
4936 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4937 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4938 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4939 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4940 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4941 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4942 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4943 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4944 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4945 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4946 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4947 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4948 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4949 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4950 // AVX 256-bit support
4951 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4952 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4953 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4954 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4955 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4956 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4957};
4958
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4959static const uint16_t ReplaceableInstrsAVX2[][3] = {
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4960 //PackedSingle PackedDouble PackedInt
Craig Topperf87a2be2011-11-09 09:37:21 +0000[diff] [blame]4961 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4962 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4963 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4964 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4965 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4966 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4967 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
Craig Topper12b72de2011-11-29 05:37:58 +0000[diff] [blame]4968 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4969 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4970 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4971 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4972 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4973 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4974 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4975};
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4976
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4977// FIXME: Some shuffle and unpack instructions have equivalents in different
4978// domains, but they require a bit more work than just switching opcodes.
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4979
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4980static const uint16_t *lookup(unsigned opcode, unsigned domain) {
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4981 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4982 if (ReplaceableInstrs[i][domain-1] == opcode)
4983 return ReplaceableInstrs[i];
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4984 return 0;
4985}
4986
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4987static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4988 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4989 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4990 return ReplaceableInstrsAVX2[i];
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4991 return 0;
4992}
4993
4994std::pair<uint16_t, uint16_t>
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]4995X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4996 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4997 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4998 uint16_t validDomains = 0;
4999 if (domain && lookup(MI->getOpcode(), domain))
5000 validDomains = 0xe;
5001 else if (domain && lookupAVX2(MI->getOpcode(), domain))
5002 validDomains = hasAVX2 ? 0xe : 0x6;
5003 return std::make_pair(domain, validDomains);
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5004}
5005
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]5006void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5007 assert(Domain>0 && Domain<4 && "Invalid execution domain");
5008 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5009 assert(dom && "Not an SSE instruction");
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]5010 const uint16_t *table = lookup(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]5011 if (!table) { // try the other table
5012 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
5013 "256-bit vector operations only available in AVX2");
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]5014 table = lookupAVX2(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]5015 }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]5016 assert(table && "Cannot change domain");
5017 MI->setDesc(get(table[Domain-1]));
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]5018}
Chris Lattner6a5e7062010-04-26 23:37:21 +0000[diff] [blame]5019
5020/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
5021void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
5022 NopInst.setOpcode(X86::NOOP);
5023}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5024
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]5025bool X86InstrInfo::isHighLatencyDef(int opc) const {
5026 switch (opc) {
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5027 default: return false;
5028 case X86::DIVSDrm:
5029 case X86::DIVSDrm_Int:
5030 case X86::DIVSDrr:
5031 case X86::DIVSDrr_Int:
5032 case X86::DIVSSrm:
5033 case X86::DIVSSrm_Int:
5034 case X86::DIVSSrr:
5035 case X86::DIVSSrr_Int:
5036 case X86::SQRTPDm:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5037 case X86::SQRTPDr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5038 case X86::SQRTPSm:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5039 case X86::SQRTPSr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5040 case X86::SQRTSDm:
5041 case X86::SQRTSDm_Int:
5042 case X86::SQRTSDr:
5043 case X86::SQRTSDr_Int:
5044 case X86::SQRTSSm:
5045 case X86::SQRTSSm_Int:
5046 case X86::SQRTSSr:
5047 case X86::SQRTSSr_Int:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5048 // AVX instructions with high latency
5049 case X86::VDIVSDrm:
5050 case X86::VDIVSDrm_Int:
5051 case X86::VDIVSDrr:
5052 case X86::VDIVSDrr_Int:
5053 case X86::VDIVSSrm:
5054 case X86::VDIVSSrm_Int:
5055 case X86::VDIVSSrr:
5056 case X86::VDIVSSrr_Int:
5057 case X86::VSQRTPDm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5058 case X86::VSQRTPDr:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5059 case X86::VSQRTPSm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5060 case X86::VSQRTPSr:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]5061 case X86::VSQRTSDm:
5062 case X86::VSQRTSDm_Int:
5063 case X86::VSQRTSDr:
5064 case X86::VSQRTSSm:
5065 case X86::VSQRTSSm_Int:
5066 case X86::VSQRTSSr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]5067 return true;
5068 }
5069}
5070
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]5071bool X86InstrInfo::
5072hasHighOperandLatency(const InstrItineraryData *ItinData,
5073 const MachineRegisterInfo *MRI,
5074 const MachineInstr *DefMI, unsigned DefIdx,
5075 const MachineInstr *UseMI, unsigned UseIdx) const {
5076 return isHighLatencyDef(DefMI->getOpcode());
5077}
5078
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5079namespace {
5080 /// CGBR - Create Global Base Reg pass. This initializes the PIC
5081 /// global base register for x86-32.
5082 struct CGBR : public MachineFunctionPass {
5083 static char ID;
Owen Andersona7aed182010-08-06 18:33:48 +0000[diff] [blame]5084 CGBR() : MachineFunctionPass(ID) {}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5085
5086 virtual bool runOnMachineFunction(MachineFunction &MF) {
5087 const X86TargetMachine *TM =
5088 static_cast<const X86TargetMachine *>(&MF.getTarget());
5089
5090 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
5091 "X86-64 PIC uses RIP relative addressing");
5092
5093 // Only emit a global base reg in PIC mode.
5094 if (TM->getRelocationModel() != Reloc::PIC_)
5095 return false;
5096
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]5097 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
5098 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5099
5100 // If we didn't need a GlobalBaseReg, don't insert code.
5101 if (GlobalBaseReg == 0)
5102 return false;
5103
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5104 // Insert the set of GlobalBaseReg into the first MBB of the function
5105 MachineBasicBlock &FirstMBB = MF.front();
5106 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
5107 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
5108 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5109 const X86InstrInfo *TII = TM->getInstrInfo();
5110
5111 unsigned PC;
5112 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]5113 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5114 else
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]5115 PC = GlobalBaseReg;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]5116
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5117 // Operand of MovePCtoStack is completely ignored by asm printer. It's
5118 // only used in JIT code emission as displacement to pc.
5119 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]5120
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5121 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
5122 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
5123 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]5124 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
5125 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
5126 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
5127 X86II::MO_GOT_ABSOLUTE_ADDRESS);
5128 }
5129
5130 return true;
5131 }
5132
5133 virtual const char *getPassName() const {
5134 return "X86 PIC Global Base Reg Initialization";
5135 }
5136
5137 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
5138 AU.setPreservesCFG();
5139 MachineFunctionPass::getAnalysisUsage(AU);
5140 }
5141 };
5142}
5143
5144char CGBR::ID = 0;
5145FunctionPass*
5146llvm::createGlobalBaseRegPass() { return new CGBR(); }
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]5147
5148namespace {
5149 struct LDTLSCleanup : public MachineFunctionPass {
5150 static char ID;
5151 LDTLSCleanup() : MachineFunctionPass(ID) {}
5152
5153 virtual bool runOnMachineFunction(MachineFunction &MF) {
5154 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
5155 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
5156 // No point folding accesses if there isn't at least two.
5157 return false;
5158 }
5159
5160 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
5161 return VisitNode(DT->getRootNode(), 0);
5162 }
5163
5164 // Visit the dominator subtree rooted at Node in pre-order.
5165 // If TLSBaseAddrReg is non-null, then use that to replace any
5166 // TLS_base_addr instructions. Otherwise, create the register
5167 // when the first such instruction is seen, and then use it
5168 // as we encounter more instructions.
5169 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
5170 MachineBasicBlock *BB = Node->getBlock();
5171 bool Changed = false;
5172
5173 // Traverse the current block.
5174 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
5175 ++I) {
5176 switch (I->getOpcode()) {
5177 case X86::TLS_base_addr32:
5178 case X86::TLS_base_addr64:
5179 if (TLSBaseAddrReg)
5180 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
5181 else
5182 I = SetRegister(I, &TLSBaseAddrReg);
5183 Changed = true;
5184 break;
5185 default:
5186 break;
5187 }
5188 }
5189
5190 // Visit the children of this block in the dominator tree.
5191 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
5192 I != E; ++I) {
5193 Changed |= VisitNode(*I, TLSBaseAddrReg);
5194 }
5195
5196 return Changed;
5197 }
5198
5199 // Replace the TLS_base_addr instruction I with a copy from
5200 // TLSBaseAddrReg, returning the new instruction.
5201 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
5202 unsigned TLSBaseAddrReg) {
5203 MachineFunction *MF = I->getParent()->getParent();
5204 const X86TargetMachine *TM =
5205 static_cast<const X86TargetMachine *>(&MF->getTarget());
5206 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5207 const X86InstrInfo *TII = TM->getInstrInfo();
5208
5209 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
5210 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
5211 TII->get(TargetOpcode::COPY),
5212 is64Bit ? X86::RAX : X86::EAX)
5213 .addReg(TLSBaseAddrReg);
5214
5215 // Erase the TLS_base_addr instruction.
5216 I->eraseFromParent();
5217
5218 return Copy;
5219 }
5220
5221 // Create a virtal register in *TLSBaseAddrReg, and populate it by
5222 // inserting a copy instruction after I. Returns the new instruction.
5223 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
5224 MachineFunction *MF = I->getParent()->getParent();
5225 const X86TargetMachine *TM =
5226 static_cast<const X86TargetMachine *>(&MF->getTarget());
5227 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5228 const X86InstrInfo *TII = TM->getInstrInfo();
5229
5230 // Create a virtual register for the TLS base address.
5231 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5232 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
5233 ? &X86::GR64RegClass
5234 : &X86::GR32RegClass);
5235
5236 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
5237 MachineInstr *Next = I->getNextNode();
5238 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
5239 TII->get(TargetOpcode::COPY),
5240 *TLSBaseAddrReg)
5241 .addReg(is64Bit ? X86::RAX : X86::EAX);
5242
5243 return Copy;
5244 }
5245
5246 virtual const char *getPassName() const {
5247 return "Local Dynamic TLS Access Clean-up";
5248 }
5249
5250 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
5251 AU.setPreservesCFG();
5252 AU.addRequired<MachineDominatorTree>();
5253 MachineFunctionPass::getAnalysisUsage(AU);
5254 }
5255 };
5256}
5257
5258char LDTLSCleanup::ID = 0;
5259FunctionPass*
5260llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }