Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / 52ca447fdd10d2c9411fe4bd482cc52240e35d2d / . / llvm / lib / Target / X86 / X86InstrInfo.cpp
blob: 5ed8604a3990a219d74cb0b66d70530ca6de44d8 [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,
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]84 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]85};
86
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]87struct X86OpTblEntry {
88 uint16_t RegOp;
89 uint16_t MemOp;
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]90 uint16_t Flags;
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]91};
92
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]93X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
Evan Cheng703a0fb2011-07-01 17:57:27 +0000[diff] [blame]94 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
95 ? X86::ADJCALLSTACKDOWN64
96 : X86::ADJCALLSTACKDOWN32),
97 (tm.getSubtarget<X86Subtarget>().is64Bit()
98 ? X86::ADJCALLSTACKUP64
99 : X86::ADJCALLSTACKUP32)),
Evan Cheng11b0a5d2006-09-08 06:48:29 +0000[diff] [blame]100 TM(tm), RI(tm, *this) {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]101
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]102 static const X86OpTblEntry OpTbl2Addr[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]103 { X86::ADC32ri, X86::ADC32mi, 0 },
104 { X86::ADC32ri8, X86::ADC32mi8, 0 },
105 { X86::ADC32rr, X86::ADC32mr, 0 },
106 { X86::ADC64ri32, X86::ADC64mi32, 0 },
107 { X86::ADC64ri8, X86::ADC64mi8, 0 },
108 { X86::ADC64rr, X86::ADC64mr, 0 },
109 { X86::ADD16ri, X86::ADD16mi, 0 },
110 { X86::ADD16ri8, X86::ADD16mi8, 0 },
111 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
112 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
113 { X86::ADD16rr, X86::ADD16mr, 0 },
114 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
115 { X86::ADD32ri, X86::ADD32mi, 0 },
116 { X86::ADD32ri8, X86::ADD32mi8, 0 },
117 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
118 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
119 { X86::ADD32rr, X86::ADD32mr, 0 },
120 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
121 { X86::ADD64ri32, X86::ADD64mi32, 0 },
122 { X86::ADD64ri8, X86::ADD64mi8, 0 },
123 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
124 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
125 { X86::ADD64rr, X86::ADD64mr, 0 },
126 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
127 { X86::ADD8ri, X86::ADD8mi, 0 },
128 { X86::ADD8rr, X86::ADD8mr, 0 },
129 { X86::AND16ri, X86::AND16mi, 0 },
130 { X86::AND16ri8, X86::AND16mi8, 0 },
131 { X86::AND16rr, X86::AND16mr, 0 },
132 { X86::AND32ri, X86::AND32mi, 0 },
133 { X86::AND32ri8, X86::AND32mi8, 0 },
134 { X86::AND32rr, X86::AND32mr, 0 },
135 { X86::AND64ri32, X86::AND64mi32, 0 },
136 { X86::AND64ri8, X86::AND64mi8, 0 },
137 { X86::AND64rr, X86::AND64mr, 0 },
138 { X86::AND8ri, X86::AND8mi, 0 },
139 { X86::AND8rr, X86::AND8mr, 0 },
140 { X86::DEC16r, X86::DEC16m, 0 },
141 { X86::DEC32r, X86::DEC32m, 0 },
142 { X86::DEC64_16r, X86::DEC64_16m, 0 },
143 { X86::DEC64_32r, X86::DEC64_32m, 0 },
144 { X86::DEC64r, X86::DEC64m, 0 },
145 { X86::DEC8r, X86::DEC8m, 0 },
146 { X86::INC16r, X86::INC16m, 0 },
147 { X86::INC32r, X86::INC32m, 0 },
148 { X86::INC64_16r, X86::INC64_16m, 0 },
149 { X86::INC64_32r, X86::INC64_32m, 0 },
150 { X86::INC64r, X86::INC64m, 0 },
151 { X86::INC8r, X86::INC8m, 0 },
152 { X86::NEG16r, X86::NEG16m, 0 },
153 { X86::NEG32r, X86::NEG32m, 0 },
154 { X86::NEG64r, X86::NEG64m, 0 },
155 { X86::NEG8r, X86::NEG8m, 0 },
156 { X86::NOT16r, X86::NOT16m, 0 },
157 { X86::NOT32r, X86::NOT32m, 0 },
158 { X86::NOT64r, X86::NOT64m, 0 },
159 { X86::NOT8r, X86::NOT8m, 0 },
160 { X86::OR16ri, X86::OR16mi, 0 },
161 { X86::OR16ri8, X86::OR16mi8, 0 },
162 { X86::OR16rr, X86::OR16mr, 0 },
163 { X86::OR32ri, X86::OR32mi, 0 },
164 { X86::OR32ri8, X86::OR32mi8, 0 },
165 { X86::OR32rr, X86::OR32mr, 0 },
166 { X86::OR64ri32, X86::OR64mi32, 0 },
167 { X86::OR64ri8, X86::OR64mi8, 0 },
168 { X86::OR64rr, X86::OR64mr, 0 },
169 { X86::OR8ri, X86::OR8mi, 0 },
170 { X86::OR8rr, X86::OR8mr, 0 },
171 { X86::ROL16r1, X86::ROL16m1, 0 },
172 { X86::ROL16rCL, X86::ROL16mCL, 0 },
173 { X86::ROL16ri, X86::ROL16mi, 0 },
174 { X86::ROL32r1, X86::ROL32m1, 0 },
175 { X86::ROL32rCL, X86::ROL32mCL, 0 },
176 { X86::ROL32ri, X86::ROL32mi, 0 },
177 { X86::ROL64r1, X86::ROL64m1, 0 },
178 { X86::ROL64rCL, X86::ROL64mCL, 0 },
179 { X86::ROL64ri, X86::ROL64mi, 0 },
180 { X86::ROL8r1, X86::ROL8m1, 0 },
181 { X86::ROL8rCL, X86::ROL8mCL, 0 },
182 { X86::ROL8ri, X86::ROL8mi, 0 },
183 { X86::ROR16r1, X86::ROR16m1, 0 },
184 { X86::ROR16rCL, X86::ROR16mCL, 0 },
185 { X86::ROR16ri, X86::ROR16mi, 0 },
186 { X86::ROR32r1, X86::ROR32m1, 0 },
187 { X86::ROR32rCL, X86::ROR32mCL, 0 },
188 { X86::ROR32ri, X86::ROR32mi, 0 },
189 { X86::ROR64r1, X86::ROR64m1, 0 },
190 { X86::ROR64rCL, X86::ROR64mCL, 0 },
191 { X86::ROR64ri, X86::ROR64mi, 0 },
192 { X86::ROR8r1, X86::ROR8m1, 0 },
193 { X86::ROR8rCL, X86::ROR8mCL, 0 },
194 { X86::ROR8ri, X86::ROR8mi, 0 },
195 { X86::SAR16r1, X86::SAR16m1, 0 },
196 { X86::SAR16rCL, X86::SAR16mCL, 0 },
197 { X86::SAR16ri, X86::SAR16mi, 0 },
198 { X86::SAR32r1, X86::SAR32m1, 0 },
199 { X86::SAR32rCL, X86::SAR32mCL, 0 },
200 { X86::SAR32ri, X86::SAR32mi, 0 },
201 { X86::SAR64r1, X86::SAR64m1, 0 },
202 { X86::SAR64rCL, X86::SAR64mCL, 0 },
203 { X86::SAR64ri, X86::SAR64mi, 0 },
204 { X86::SAR8r1, X86::SAR8m1, 0 },
205 { X86::SAR8rCL, X86::SAR8mCL, 0 },
206 { X86::SAR8ri, X86::SAR8mi, 0 },
207 { X86::SBB32ri, X86::SBB32mi, 0 },
208 { X86::SBB32ri8, X86::SBB32mi8, 0 },
209 { X86::SBB32rr, X86::SBB32mr, 0 },
210 { X86::SBB64ri32, X86::SBB64mi32, 0 },
211 { X86::SBB64ri8, X86::SBB64mi8, 0 },
212 { X86::SBB64rr, X86::SBB64mr, 0 },
213 { X86::SHL16rCL, X86::SHL16mCL, 0 },
214 { X86::SHL16ri, X86::SHL16mi, 0 },
215 { X86::SHL32rCL, X86::SHL32mCL, 0 },
216 { X86::SHL32ri, X86::SHL32mi, 0 },
217 { X86::SHL64rCL, X86::SHL64mCL, 0 },
218 { X86::SHL64ri, X86::SHL64mi, 0 },
219 { X86::SHL8rCL, X86::SHL8mCL, 0 },
220 { X86::SHL8ri, X86::SHL8mi, 0 },
221 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
222 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
223 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
224 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
225 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
226 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
227 { X86::SHR16r1, X86::SHR16m1, 0 },
228 { X86::SHR16rCL, X86::SHR16mCL, 0 },
229 { X86::SHR16ri, X86::SHR16mi, 0 },
230 { X86::SHR32r1, X86::SHR32m1, 0 },
231 { X86::SHR32rCL, X86::SHR32mCL, 0 },
232 { X86::SHR32ri, X86::SHR32mi, 0 },
233 { X86::SHR64r1, X86::SHR64m1, 0 },
234 { X86::SHR64rCL, X86::SHR64mCL, 0 },
235 { X86::SHR64ri, X86::SHR64mi, 0 },
236 { X86::SHR8r1, X86::SHR8m1, 0 },
237 { X86::SHR8rCL, X86::SHR8mCL, 0 },
238 { X86::SHR8ri, X86::SHR8mi, 0 },
239 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
240 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
241 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
242 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
243 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
244 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
245 { X86::SUB16ri, X86::SUB16mi, 0 },
246 { X86::SUB16ri8, X86::SUB16mi8, 0 },
247 { X86::SUB16rr, X86::SUB16mr, 0 },
248 { X86::SUB32ri, X86::SUB32mi, 0 },
249 { X86::SUB32ri8, X86::SUB32mi8, 0 },
250 { X86::SUB32rr, X86::SUB32mr, 0 },
251 { X86::SUB64ri32, X86::SUB64mi32, 0 },
252 { X86::SUB64ri8, X86::SUB64mi8, 0 },
253 { X86::SUB64rr, X86::SUB64mr, 0 },
254 { X86::SUB8ri, X86::SUB8mi, 0 },
255 { X86::SUB8rr, X86::SUB8mr, 0 },
256 { X86::XOR16ri, X86::XOR16mi, 0 },
257 { X86::XOR16ri8, X86::XOR16mi8, 0 },
258 { X86::XOR16rr, X86::XOR16mr, 0 },
259 { X86::XOR32ri, X86::XOR32mi, 0 },
260 { X86::XOR32ri8, X86::XOR32mi8, 0 },
261 { X86::XOR32rr, X86::XOR32mr, 0 },
262 { X86::XOR64ri32, X86::XOR64mi32, 0 },
263 { X86::XOR64ri8, X86::XOR64mi8, 0 },
264 { X86::XOR64rr, X86::XOR64mr, 0 },
265 { X86::XOR8ri, X86::XOR8mi, 0 },
266 { X86::XOR8rr, X86::XOR8mr, 0 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]267 };
268
269 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]270 unsigned RegOp = OpTbl2Addr[i].RegOp;
271 unsigned MemOp = OpTbl2Addr[i].MemOp;
272 unsigned Flags = OpTbl2Addr[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]273 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
274 RegOp, MemOp,
275 // Index 0, folded load and store, no alignment requirement.
276 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]277 }
278
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]279 static const X86OpTblEntry OpTbl0[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]280 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
281 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
282 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
283 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
284 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]285 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
286 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
287 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
288 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
289 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
290 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
291 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
292 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
293 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
294 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
295 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
296 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
297 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
298 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
299 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
Craig Topperd09a9af2012-12-26 01:47:12 +0000[diff] [blame]300 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]301 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
302 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]303 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
304 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
305 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
306 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
307 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
308 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
309 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
310 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
311 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
312 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
313 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
314 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
315 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
316 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
317 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
318 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
319 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
320 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
321 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
322 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
323 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
324 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]325 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
326 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
327 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
328 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
329 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
330 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]331 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
332 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
333 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
334 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
335 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
336 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
337 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
338 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
339 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
340 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
341 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
342 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
343 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
344 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
345 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
346 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
347 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
348 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
349 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
350 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
351 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
352 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
353 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
354 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
355 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]356 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
357 // AVX 128-bit versions of foldable instructions
Craig Topperd09a9af2012-12-26 01:47:12 +0000[diff] [blame]358 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]359 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
360 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]361 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]362 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
363 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
366 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
367 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
368 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
369 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
370 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
371 // AVX 256-bit foldable instructions
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]372 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]373 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
374 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
377 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]378 };
379
380 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]381 unsigned RegOp = OpTbl0[i].RegOp;
382 unsigned MemOp = OpTbl0[i].MemOp;
383 unsigned Flags = OpTbl0[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]384 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
385 RegOp, MemOp, TB_INDEX_0 | Flags);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]386 }
387
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]388 static const X86OpTblEntry OpTbl1[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]389 { X86::CMP16rr, X86::CMP16rm, 0 },
390 { X86::CMP32rr, X86::CMP32rm, 0 },
391 { X86::CMP64rr, X86::CMP64rm, 0 },
392 { X86::CMP8rr, X86::CMP8rm, 0 },
393 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
394 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
395 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
396 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
397 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
398 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
399 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
400 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
401 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
402 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
403 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
404 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]405 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
406 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
407 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
408 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
409 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
410 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
411 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
412 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]413 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
414 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]415 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
416 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]417 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
418 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
419 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
420 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
421 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
422 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
423 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
424 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]425 { X86::MOV16rr, X86::MOV16rm, 0 },
426 { X86::MOV32rr, X86::MOV32rm, 0 },
427 { X86::MOV64rr, X86::MOV64rm, 0 },
428 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
429 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
430 { X86::MOV8rr, X86::MOV8rm, 0 },
431 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
432 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]433 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
434 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
435 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
436 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]437 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
438 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
439 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
440 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
441 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
442 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
443 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
444 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
445 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
446 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]447 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
448 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
449 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
450 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
451 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
452 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
453 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]454 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
455 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
456 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]457 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
458 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
459 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
460 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
461 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
462 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
463 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
464 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
465 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
466 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]467 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]468 { X86::SQRTSDr, X86::SQRTSDm, 0 },
469 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
470 { X86::SQRTSSr, X86::SQRTSSm, 0 },
471 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
472 { X86::TEST16rr, X86::TEST16rm, 0 },
473 { X86::TEST32rr, X86::TEST32rm, 0 },
474 { X86::TEST64rr, X86::TEST64rm, 0 },
475 { X86::TEST8rr, X86::TEST8rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]476 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]477 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
478 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]479 // AVX 128-bit versions of foldable instructions
480 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
481 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]482 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
483 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]484 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
485 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
Pete Cooper8bbce762012-06-14 22:12:58 +0000[diff] [blame]486 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]487 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
488 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
489 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
490 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
491 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
492 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
493 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
494 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
495 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]496 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
497 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
498 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
499 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
500 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
501 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
502 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
503 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
504 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
505 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
506 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
507 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
Craig Topperb2922162012-12-26 02:14:19 +0000[diff] [blame]508 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]509 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
510 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
511 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
512 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]513 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
514 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
515 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
516 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
517 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
518 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
519 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
520 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
521 { X86::VRCPPSr, X86::VRCPPSm, 0 },
522 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, 0 },
523 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
524 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, 0 },
525 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]526 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]527 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]528 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]529 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
530
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]531 // AVX 256-bit foldable instructions
532 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
533 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]534 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]535 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]536 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]537 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
538 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]539
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]540 // AVX2 foldable instructions
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]541 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
542 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
543 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
544 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
545 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
546 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
547 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
548 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, 0 },
549 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]550 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]551 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]552 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
553 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
Michael Liao2de86af2012-09-26 08:24:51 +0000[diff] [blame]554
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]555 // BMI/BMI2/LZCNT/POPCNT foldable instructions
556 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
557 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
558 { X86::BLSI32rr, X86::BLSI32rm, 0 },
559 { X86::BLSI64rr, X86::BLSI64rm, 0 },
560 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
561 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
562 { X86::BLSR32rr, X86::BLSR32rm, 0 },
563 { X86::BLSR64rr, X86::BLSR64rm, 0 },
564 { X86::BZHI32rr, X86::BZHI32rm, 0 },
565 { X86::BZHI64rr, X86::BZHI64rm, 0 },
566 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
567 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
568 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
569 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
570 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
571 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
Michael Liao2de86af2012-09-26 08:24:51 +0000[diff] [blame]572 { X86::RORX32ri, X86::RORX32mi, 0 },
573 { X86::RORX64ri, X86::RORX64mi, 0 },
Michael Liao2b425e12012-09-26 08:26:25 +0000[diff] [blame]574 { X86::SARX32rr, X86::SARX32rm, 0 },
575 { X86::SARX64rr, X86::SARX64rm, 0 },
576 { X86::SHRX32rr, X86::SHRX32rm, 0 },
577 { X86::SHRX64rr, X86::SHRX64rm, 0 },
578 { X86::SHLX32rr, X86::SHLX32rm, 0 },
579 { X86::SHLX64rr, X86::SHLX64rm, 0 },
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]580 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
581 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
582 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]583 };
584
585 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]586 unsigned RegOp = OpTbl1[i].RegOp;
587 unsigned MemOp = OpTbl1[i].MemOp;
588 unsigned Flags = OpTbl1[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]589 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
590 RegOp, MemOp,
591 // Index 1, folded load
592 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]593 }
594
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]595 static const X86OpTblEntry OpTbl2[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]596 { X86::ADC32rr, X86::ADC32rm, 0 },
597 { X86::ADC64rr, X86::ADC64rm, 0 },
598 { X86::ADD16rr, X86::ADD16rm, 0 },
599 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
600 { X86::ADD32rr, X86::ADD32rm, 0 },
601 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
602 { X86::ADD64rr, X86::ADD64rm, 0 },
603 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
604 { X86::ADD8rr, X86::ADD8rm, 0 },
605 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
606 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
607 { X86::ADDSDrr, X86::ADDSDrm, 0 },
608 { X86::ADDSSrr, X86::ADDSSrm, 0 },
609 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
610 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
611 { X86::AND16rr, X86::AND16rm, 0 },
612 { X86::AND32rr, X86::AND32rm, 0 },
613 { X86::AND64rr, X86::AND64rm, 0 },
614 { X86::AND8rr, X86::AND8rm, 0 },
615 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
616 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
617 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
618 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]619 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
620 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
621 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
622 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]623 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
624 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
625 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
626 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
627 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
628 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
629 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
630 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
631 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
632 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
633 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
634 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
635 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
636 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
637 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
638 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
639 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
640 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
641 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
642 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
643 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
644 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
645 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
646 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
647 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
648 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
649 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
650 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
651 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
652 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
653 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
654 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
655 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
656 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
657 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
658 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
659 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
660 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
661 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
662 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
663 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
664 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
665 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
666 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
667 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
668 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
669 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
670 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
671 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
672 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
673 { X86::CMPSDrr, X86::CMPSDrm, 0 },
674 { X86::CMPSSrr, X86::CMPSSrm, 0 },
675 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
676 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
677 { X86::DIVSDrr, X86::DIVSDrm, 0 },
678 { X86::DIVSSrr, X86::DIVSSrm, 0 },
679 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
680 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
681 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
682 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
683 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
684 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
685 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
686 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
687 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
688 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
689 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
690 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
691 { X86::IMUL16rr, X86::IMUL16rm, 0 },
692 { X86::IMUL32rr, X86::IMUL32rm, 0 },
693 { X86::IMUL64rr, X86::IMUL64rm, 0 },
694 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
695 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
Manman Ren959acb12012-08-13 18:29:41 +0000[diff] [blame]696 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
697 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
698 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
699 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
700 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
701 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]702 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]703 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]704 { X86::MAXSDrr, X86::MAXSDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]705 { X86::MAXSSrr, X86::MAXSSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]706 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]707 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]708 { X86::MINSDrr, X86::MINSDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]709 { X86::MINSSrr, X86::MINSSrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]710 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]711 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
712 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
713 { X86::MULSDrr, X86::MULSDrm, 0 },
714 { X86::MULSSrr, X86::MULSSrm, 0 },
715 { X86::OR16rr, X86::OR16rm, 0 },
716 { X86::OR32rr, X86::OR32rm, 0 },
717 { X86::OR64rr, X86::OR64rm, 0 },
718 { X86::OR8rr, X86::OR8rm, 0 },
719 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
720 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
721 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
722 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]723 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]724 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
725 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
726 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
727 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
728 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
729 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]730 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
731 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]732 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]733 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]734 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
735 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
736 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
737 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]738 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]739 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
740 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]741 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]742 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
743 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
744 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]745 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]746 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]747 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
748 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]749 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]750 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]751 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]752 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]753 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]754 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]755 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
756 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
757 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
758 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
759 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
Benjamin Kramer4669d182012-12-21 14:04:55 +0000[diff] [blame]760 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
761 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
762 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
763 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
764 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
765 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
766 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
767 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]768 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]769 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]770 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
771 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
772 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
773 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
774 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
775 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
776 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]777 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
778 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
779 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
780 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]781 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
782 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
783 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
784 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
785 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
786 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
787 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
788 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
789 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
790 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
791 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
792 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
793 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
794 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
795 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
796 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
797 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
798 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
799 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
800 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
801 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
802 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
803 { X86::SBB32rr, X86::SBB32rm, 0 },
804 { X86::SBB64rr, X86::SBB64rm, 0 },
805 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
806 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
807 { X86::SUB16rr, X86::SUB16rm, 0 },
808 { X86::SUB32rr, X86::SUB32rm, 0 },
809 { X86::SUB64rr, X86::SUB64rm, 0 },
810 { X86::SUB8rr, X86::SUB8rm, 0 },
811 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
812 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
813 { X86::SUBSDrr, X86::SUBSDrm, 0 },
814 { X86::SUBSSrr, X86::SUBSSrm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]815 // FIXME: TEST*rr -> swapped operand of TEST*mr.
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]816 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
817 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
818 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
819 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
820 { X86::XOR16rr, X86::XOR16rm, 0 },
821 { X86::XOR32rr, X86::XOR32rm, 0 },
822 { X86::XOR64rr, X86::XOR64rm, 0 },
823 { X86::XOR8rr, X86::XOR8rm, 0 },
824 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]825 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
826 // AVX 128-bit versions of foldable instructions
827 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
828 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
829 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
830 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
831 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
832 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
833 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
834 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
835 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
836 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
Craig Toppercaef1c52012-12-26 00:35:47 +0000[diff] [blame]837 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
838 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]839 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
840 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]841 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
842 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
843 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]844 { X86::VADDPDrr, X86::VADDPDrm, 0 },
845 { X86::VADDPSrr, X86::VADDPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]846 { X86::VADDSDrr, X86::VADDSDrm, 0 },
847 { X86::VADDSSrr, X86::VADDSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]848 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
849 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
850 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
851 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
852 { X86::VANDPDrr, X86::VANDPDrm, 0 },
853 { X86::VANDPSrr, X86::VANDPSrm, 0 },
854 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
855 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
856 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
857 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
858 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
859 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]860 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
861 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]862 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
863 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]864 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
865 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
866 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
867 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
868 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
869 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
870 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
871 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
872 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
873 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]874 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
875 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
876 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
877 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]878 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
879 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]880 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]881 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]882 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]883 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]884 { X86::VMINPDrr, X86::VMINPDrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]885 { X86::VMINPSrr, X86::VMINPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]886 { X86::VMINSDrr, X86::VMINSDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]887 { X86::VMINSSrr, X86::VMINSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]888 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
889 { X86::VMULPDrr, X86::VMULPDrm, 0 },
890 { X86::VMULPSrr, X86::VMULPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]891 { X86::VMULSDrr, X86::VMULSDrm, 0 },
892 { X86::VMULSSrr, X86::VMULSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]893 { X86::VORPDrr, X86::VORPDrm, 0 },
894 { X86::VORPSrr, X86::VORPSrm, 0 },
895 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
896 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
897 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
898 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
899 { X86::VPADDBrr, X86::VPADDBrm, 0 },
900 { X86::VPADDDrr, X86::VPADDDrm, 0 },
901 { X86::VPADDQrr, X86::VPADDQrm, 0 },
902 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
903 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
904 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
905 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
906 { X86::VPADDWrr, X86::VPADDWrm, 0 },
907 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
908 { X86::VPANDNrr, X86::VPANDNrm, 0 },
909 { X86::VPANDrr, X86::VPANDrm, 0 },
910 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
911 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
912 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
913 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
914 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
915 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
916 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
917 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
918 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
919 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
920 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
921 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
922 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
923 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
924 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
925 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
926 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
927 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
928 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
929 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
930 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
931 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
932 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
933 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
934 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
935 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
936 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
937 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
938 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
939 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
940 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
941 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
942 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
943 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
944 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
945 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
946 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
947 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
948 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
949 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
950 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
951 { X86::VPORrr, X86::VPORrm, 0 },
952 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
953 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
954 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
955 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
956 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
957 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
958 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
959 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
960 { X86::VPSRADrr, X86::VPSRADrm, 0 },
961 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
962 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
963 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
964 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
965 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
966 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
967 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
968 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
969 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
970 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
971 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
972 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
973 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
974 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
975 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
976 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
977 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
978 { X86::VPXORrr, X86::VPXORrm, 0 },
979 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
980 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
981 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
982 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]983 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
984 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]985 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
986 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
987 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
988 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
989 { X86::VXORPDrr, X86::VXORPDrm, 0 },
990 { X86::VXORPSrr, X86::VXORPSrm, 0 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]991 // AVX 256-bit foldable instructions
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]992 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
993 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
994 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
995 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
996 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
997 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
998 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
999 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1000 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1001 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1002 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1003 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1004 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1005 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1006 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1007 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1008 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1009 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1010 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1011 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1012 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1013 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1014 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1015 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1016 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1017 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1018 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1019 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1020 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1021 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1022 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1023 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1024 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1025 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1026 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1027 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1028 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1029 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1030 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1031 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1032 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1033 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1034 // AVX2 foldable instructions
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1035 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1036 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1037 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1038 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1039 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1040 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1041 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1042 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1043 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1044 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1045 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1046 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1047 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1048 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1049 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1050 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1051 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1052 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1053 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1054 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1055 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1056 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1057 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1058 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1059 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1060 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1061 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1062 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1063 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1064 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1065 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1066 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
1067 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1068 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
1069 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1070 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1071 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1072 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1073 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1074 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1075 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1076 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1077 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1078 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1079 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1080 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1081 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1082 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1083 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1084 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1085 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1086 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1087 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1088 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1089 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1090 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1091 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1092 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1093 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1094 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1095 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1096 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1097 { X86::VPORYrr, X86::VPORYrm, 0 },
1098 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1099 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1100 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1101 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1102 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1103 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1104 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1105 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1106 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1107 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1108 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1109 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1110 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1111 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1112 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1113 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1114 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1115 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1116 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1117 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1118 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1119 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1120 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1121 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1122 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1123 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1124 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1125 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1126 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1127 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1128 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1129 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1130 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1131 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1132 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1133 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1134 { X86::VPXORYrr, X86::VPXORYrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1135 // FIXME: add AVX 256-bit foldable instructions
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1136
1137 // FMA4 foldable patterns
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1138 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, 0 },
1139 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1140 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1141 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1142 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1143 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1144 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, 0 },
1145 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1146 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1147 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1148 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1149 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1150 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, 0 },
1151 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1152 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1153 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1154 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1155 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1156 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, 0 },
1157 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1158 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1159 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1160 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1161 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1162 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1163 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1164 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1165 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1166 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1167 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1168 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1169 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
Michael Liaof9f7b552012-09-26 08:22:37 +0000[diff] [blame]1170
1171 // BMI/BMI2 foldable instructions
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]1172 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1173 { X86::ANDN64rr, X86::ANDN64rm, 0 },
Michael Liaof9f7b552012-09-26 08:22:37 +0000[diff] [blame]1174 { X86::MULX32rr, X86::MULX32rm, 0 },
1175 { X86::MULX64rr, X86::MULX64rm, 0 },
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]1176 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1177 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1178 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1179 { X86::PEXT64rr, X86::PEXT64rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1180 };
1181
1182 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]1183 unsigned RegOp = OpTbl2[i].RegOp;
1184 unsigned MemOp = OpTbl2[i].MemOp;
1185 unsigned Flags = OpTbl2[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1186 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1187 RegOp, MemOp,
1188 // Index 2, folded load
1189 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1190 }
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1191
1192 static const X86OpTblEntry OpTbl3[] = {
1193 // FMA foldable instructions
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1194 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1195 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1196 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1197 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1198 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1199 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1200 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, 0 },
1201 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1202
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1203 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1204 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1205 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1206 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1207 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1208 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1209 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1210 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1211 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1212 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1213 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1214 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1215
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1216 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1217 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1218 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1219 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1220 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1221 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1222 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, 0 },
1223 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1224
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1225 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1226 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1227 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1228 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1229 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1230 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1231 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1232 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1233 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1234 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1235 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1236 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1237
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1238 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1239 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1240 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1241 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1242 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1243 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1244 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, 0 },
1245 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1246
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1247 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1248 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1249 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1250 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1251 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1252 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1253 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1254 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1255 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1256 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1257 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1258 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1259
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1260 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1261 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1262 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1263 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1264 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1265 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1266 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, 0 },
1267 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, 0 },
Craig Topper2e127b52012-06-01 05:48:39 +0000[diff] [blame]1268
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1269 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1270 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1271 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1272 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1273 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1274 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1275 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1276 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1277 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1278 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1279 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1280 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1281
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1282 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1283 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1284 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1285 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1286 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1287 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1288 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1289 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1290 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1291 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1292 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1293 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1294
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1295 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1296 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1297 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1298 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1299 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1300 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1301 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1302 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1303 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1304 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1305 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1306 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1307
1308 // FMA4 foldable patterns
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1309 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, 0 },
1310 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1311 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1312 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1313 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1314 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1315 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, 0 },
1316 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1317 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1318 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1319 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1320 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1321 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, 0 },
1322 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1323 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1324 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1325 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1326 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1327 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, 0 },
1328 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1329 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1330 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1331 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1332 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1333 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1334 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1335 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1336 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1337 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1338 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1339 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1340 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1341 };
1342
1343 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1344 unsigned RegOp = OpTbl3[i].RegOp;
1345 unsigned MemOp = OpTbl3[i].MemOp;
1346 unsigned Flags = OpTbl3[i].Flags;
1347 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1348 RegOp, MemOp,
1349 // Index 3, folded load
1350 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1351 }
1352
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]1353}
1354
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1355void
1356X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1357 MemOp2RegOpTableType &M2RTable,
1358 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1359 if ((Flags & TB_NO_FORWARD) == 0) {
1360 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1361 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1362 }
1363 if ((Flags & TB_NO_REVERSE) == 0) {
1364 assert(!M2RTable.count(MemOp) &&
1365 "Duplicated entries in unfolding maps?");
1366 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1367 }
1368}
1369
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1370bool
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1371X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1372 unsigned &SrcReg, unsigned &DstReg,
1373 unsigned &SubIdx) const {
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1374 switch (MI.getOpcode()) {
1375 default: break;
1376 case X86::MOVSX16rr8:
1377 case X86::MOVZX16rr8:
1378 case X86::MOVSX32rr8:
1379 case X86::MOVZX32rr8:
1380 case X86::MOVSX64rr8:
Evan Chengceb5a4e2010-01-13 08:01:32 +0000[diff] [blame]1381 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1382 // It's not always legal to reference the low 8-bit of the larger
1383 // register in 32-bit mode.
1384 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1385 case X86::MOVSX32rr16:
1386 case X86::MOVZX32rr16:
1387 case X86::MOVSX64rr16:
Tim Northover04eb4232013-05-30 10:43:18 +0000[diff] [blame]1388 case X86::MOVSX64rr32: {
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1389 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1390 // Be conservative.
1391 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1392 SrcReg = MI.getOperand(1).getReg();
1393 DstReg = MI.getOperand(0).getReg();
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1394 switch (MI.getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1395 default: llvm_unreachable("Unreachable!");
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1396 case X86::MOVSX16rr8:
1397 case X86::MOVZX16rr8:
1398 case X86::MOVSX32rr8:
1399 case X86::MOVZX32rr8:
1400 case X86::MOVSX64rr8:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1401 SubIdx = X86::sub_8bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1402 break;
1403 case X86::MOVSX32rr16:
1404 case X86::MOVZX32rr16:
1405 case X86::MOVSX64rr16:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1406 SubIdx = X86::sub_16bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1407 break;
1408 case X86::MOVSX64rr32:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1409 SubIdx = X86::sub_32bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1410 break;
1411 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1412 return true;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1413 }
1414 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1415 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1416}
1417
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1418/// isFrameOperand - Return true and the FrameIndex if the specified
1419/// operand and follow operands form a reference to the stack frame.
1420bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1421 int &FrameIndex) const {
1422 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1423 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1424 MI->getOperand(Op+1).getImm() == 1 &&
1425 MI->getOperand(Op+2).getReg() == 0 &&
1426 MI->getOperand(Op+3).getImm() == 0) {
1427 FrameIndex = MI->getOperand(Op).getIndex();
1428 return true;
1429 }
1430 return false;
1431}
1432
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1433static bool isFrameLoadOpcode(int Opcode) {
1434 switch (Opcode) {
David Blaikie46a9f012012-01-20 21:51:11 +0000[diff] [blame]1435 default:
1436 return false;
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1437 case X86::MOV8rm:
1438 case X86::MOV16rm:
1439 case X86::MOV32rm:
Evan Cheng11b0a5d2006-09-08 06:48:29 +0000[diff] [blame]1440 case X86::MOV64rm:
Dale Johannesen3d7008c2007-07-04 21:07:47 +0000[diff] [blame]1441 case X86::LD_Fp64m:
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1442 case X86::MOVSSrm:
1443 case X86::MOVSDrm:
Chris Lattnerbfc2c682006-04-18 16:44:51 +0000[diff] [blame]1444 case X86::MOVAPSrm:
1445 case X86::MOVAPDrm:
Dan Gohmanbdc0f8b2009-01-09 02:40:34 +0000[diff] [blame]1446 case X86::MOVDQArm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1447 case X86::VMOVSSrm:
1448 case X86::VMOVSDrm:
1449 case X86::VMOVAPSrm:
1450 case X86::VMOVAPDrm:
1451 case X86::VMOVDQArm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1452 case X86::VMOVAPSYrm:
1453 case X86::VMOVAPDYrm:
1454 case X86::VMOVDQAYrm:
Bill Wendlinge7b2a862007-04-03 06:00:37 +0000[diff] [blame]1455 case X86::MMX_MOVD64rm:
1456 case X86::MMX_MOVQ64rm:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1457 return true;
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1458 }
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1459}
1460
1461static bool isFrameStoreOpcode(int Opcode) {
1462 switch (Opcode) {
1463 default: break;
1464 case X86::MOV8mr:
1465 case X86::MOV16mr:
1466 case X86::MOV32mr:
1467 case X86::MOV64mr:
1468 case X86::ST_FpP64m:
1469 case X86::MOVSSmr:
1470 case X86::MOVSDmr:
1471 case X86::MOVAPSmr:
1472 case X86::MOVAPDmr:
1473 case X86::MOVDQAmr:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1474 case X86::VMOVSSmr:
1475 case X86::VMOVSDmr:
1476 case X86::VMOVAPSmr:
1477 case X86::VMOVAPDmr:
1478 case X86::VMOVDQAmr:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1479 case X86::VMOVAPSYmr:
1480 case X86::VMOVAPDYmr:
1481 case X86::VMOVDQAYmr:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1482 case X86::MMX_MOVD64mr:
1483 case X86::MMX_MOVQ64mr:
1484 case X86::MMX_MOVNTQmr:
1485 return true;
1486 }
1487 return false;
1488}
1489
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1490unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1491 int &FrameIndex) const {
1492 if (isFrameLoadOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1493 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1494 return MI->getOperand(0).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1495 return 0;
1496}
1497
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1498unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1499 int &FrameIndex) const {
1500 if (isFrameLoadOpcode(MI->getOpcode())) {
1501 unsigned Reg;
1502 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1503 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1504 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1505 const MachineMemOperand *Dummy;
1506 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1507 }
1508 return 0;
1509}
1510
Dan Gohman0b273252008-11-18 19:49:32 +0000[diff] [blame]1511unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1512 int &FrameIndex) const {
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1513 if (isFrameStoreOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1514 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1515 isFrameOperand(MI, 0, FrameIndex))
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]1516 return MI->getOperand(X86::AddrNumOperands).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1517 return 0;
1518}
1519
1520unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1521 int &FrameIndex) const {
1522 if (isFrameStoreOpcode(MI->getOpcode())) {
1523 unsigned Reg;
1524 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1525 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1526 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1527 const MachineMemOperand *Dummy;
1528 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1529 }
1530 return 0;
1531}
1532
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1533/// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1534/// X86::MOVPC32r.
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1535static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
Jakob Stoklund Olesen3b9a4422012-08-08 00:40:47 +0000[diff] [blame]1536 // Don't waste compile time scanning use-def chains of physregs.
1537 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1538 return false;
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1539 bool isPICBase = false;
1540 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1541 E = MRI.def_end(); I != E; ++I) {
1542 MachineInstr *DefMI = I.getOperand().getParent();
1543 if (DefMI->getOpcode() != X86::MOVPC32r)
1544 return false;
1545 assert(!isPICBase && "More than one PIC base?");
1546 isPICBase = true;
1547 }
1548 return isPICBase;
1549}
Evan Cheng1973a462008-03-31 07:54:19 +0000[diff] [blame]1550
Bill Wendling1e117682008-05-12 20:54:26 +0000[diff] [blame]1551bool
Dan Gohmane919de52009-10-10 00:34:18 +0000[diff] [blame]1552X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1553 AliasAnalysis *AA) const {
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1554 switch (MI->getOpcode()) {
1555 default: break;
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1556 case X86::MOV8rm:
1557 case X86::MOV16rm:
1558 case X86::MOV32rm:
1559 case X86::MOV64rm:
1560 case X86::LD_Fp64m:
1561 case X86::MOVSSrm:
1562 case X86::MOVSDrm:
1563 case X86::MOVAPSrm:
1564 case X86::MOVUPSrm:
1565 case X86::MOVAPDrm:
1566 case X86::MOVDQArm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1567 case X86::MOVDQUrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1568 case X86::VMOVSSrm:
1569 case X86::VMOVSDrm:
1570 case X86::VMOVAPSrm:
1571 case X86::VMOVUPSrm:
1572 case X86::VMOVAPDrm:
1573 case X86::VMOVDQArm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1574 case X86::VMOVDQUrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1575 case X86::VMOVAPSYrm:
1576 case X86::VMOVUPSYrm:
1577 case X86::VMOVAPDYrm:
1578 case X86::VMOVDQAYrm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1579 case X86::VMOVDQUYrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1580 case X86::MMX_MOVD64rm:
1581 case X86::MMX_MOVQ64rm:
1582 case X86::FsVMOVAPSrm:
1583 case X86::FsVMOVAPDrm:
1584 case X86::FsMOVAPSrm:
1585 case X86::FsMOVAPDrm: {
1586 // Loads from constant pools are trivially rematerializable.
1587 if (MI->getOperand(1).isReg() &&
1588 MI->getOperand(2).isImm() &&
1589 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1590 MI->isInvariantLoad(AA)) {
1591 unsigned BaseReg = MI->getOperand(1).getReg();
1592 if (BaseReg == 0 || BaseReg == X86::RIP)
1593 return true;
1594 // Allow re-materialization of PIC load.
1595 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1596 return false;
1597 const MachineFunction &MF = *MI->getParent()->getParent();
1598 const MachineRegisterInfo &MRI = MF.getRegInfo();
1599 return regIsPICBase(BaseReg, MRI);
Evan Cheng94ba37f2008-02-22 09:25:47 +0000[diff] [blame]1600 }
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1601 return false;
1602 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1603
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1604 case X86::LEA32r:
1605 case X86::LEA64r: {
1606 if (MI->getOperand(2).isImm() &&
1607 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1608 !MI->getOperand(4).isReg()) {
1609 // lea fi#, lea GV, etc. are all rematerializable.
1610 if (!MI->getOperand(1).isReg())
1611 return true;
1612 unsigned BaseReg = MI->getOperand(1).getReg();
1613 if (BaseReg == 0)
1614 return true;
1615 // Allow re-materialization of lea PICBase + x.
1616 const MachineFunction &MF = *MI->getParent()->getParent();
1617 const MachineRegisterInfo &MRI = MF.getRegInfo();
1618 return regIsPICBase(BaseReg, MRI);
1619 }
1620 return false;
1621 }
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1622 }
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1623
Dan Gohmane8c1e422007-06-26 00:48:07 +0000[diff] [blame]1624 // All other instructions marked M_REMATERIALIZABLE are always trivially
1625 // rematerializable.
1626 return true;
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1627}
1628
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1629/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1630/// would clobber the EFLAGS condition register. Note the result may be
1631/// conservative. If it cannot definitely determine the safety after visiting
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1632/// a few instructions in each direction it assumes it's not safe.
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1633static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1634 MachineBasicBlock::iterator I) {
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1635 MachineBasicBlock::iterator E = MBB.end();
1636
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1637 // For compile time consideration, if we are not able to determine the
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1638 // safety after visiting 4 instructions in each direction, we will assume
1639 // it's not safe.
1640 MachineBasicBlock::iterator Iter = I;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1641 for (unsigned i = 0; Iter != E && i < 4; ++i) {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1642 bool SeenDef = false;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1643 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1644 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1645 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1646 SeenDef = true;
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1647 if (!MO.isReg())
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1648 continue;
1649 if (MO.getReg() == X86::EFLAGS) {
1650 if (MO.isUse())
1651 return false;
1652 SeenDef = true;
1653 }
1654 }
1655
1656 if (SeenDef)
1657 // This instruction defines EFLAGS, no need to look any further.
1658 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1659 ++Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1660 // Skip over DBG_VALUE.
1661 while (Iter != E && Iter->isDebugValue())
1662 ++Iter;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1663 }
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1664
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1665 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1666 // live in.
1667 if (Iter == E) {
1668 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1669 SE = MBB.succ_end(); SI != SE; ++SI)
1670 if ((*SI)->isLiveIn(X86::EFLAGS))
1671 return false;
1672 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1673 }
1674
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1675 MachineBasicBlock::iterator B = MBB.begin();
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1676 Iter = I;
1677 for (unsigned i = 0; i < 4; ++i) {
1678 // 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]1679 // EFLAGS iff EFLAGS is not live-in.
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1680 if (Iter == B)
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1681 return !MBB.isLiveIn(X86::EFLAGS);
1682
1683 --Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1684 // Skip over DBG_VALUE.
1685 while (Iter != B && Iter->isDebugValue())
1686 --Iter;
1687
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1688 bool SawKill = false;
1689 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1690 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1691 // A register mask may clobber EFLAGS, but we should still look for a
1692 // live EFLAGS def.
1693 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1694 SawKill = true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1695 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1696 if (MO.isDef()) return MO.isDead();
1697 if (MO.isKill()) SawKill = true;
1698 }
1699 }
1700
1701 if (SawKill)
1702 // This instruction kills EFLAGS and doesn't redefine it, so
1703 // there's no need to look further.
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1704 return true;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1705 }
1706
1707 // Conservative answer.
1708 return false;
1709}
1710
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1711void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1712 MachineBasicBlock::iterator I,
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1713 unsigned DestReg, unsigned SubIdx,
Evan Cheng6ad7da92009-11-14 02:55:43 +0000[diff] [blame]1714 const MachineInstr *Orig,
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1715 const TargetRegisterInfo &TRI) const {
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]1716 // MOV32r0 is implemented with a xor which clobbers condition code.
1717 // Re-materialize it as movri instructions to avoid side effects.
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1718 unsigned Opc = Orig->getOpcode();
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]1719 if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
1720 DebugLoc DL = Orig->getDebugLoc();
1721 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
1722 .addImm(0);
1723 } else {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1724 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1725 MBB.insert(I, MI);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1726 }
Evan Cheng147cb762008-04-16 23:44:44 +0000[diff] [blame]1727
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1728 MachineInstr *NewMI = prior(I);
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1729 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1730}
1731
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1732/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1733/// is not marked dead.
1734static bool hasLiveCondCodeDef(MachineInstr *MI) {
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1735 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1736 MachineOperand &MO = MI->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1737 if (MO.isReg() && MO.isDef() &&
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1738 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1739 return true;
1740 }
1741 }
1742 return false;
1743}
1744
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]1745/// getTruncatedShiftCount - check whether the shift count for a machine operand
1746/// is non-zero.
1747inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
1748 unsigned ShiftAmtOperandIdx) {
1749 // The shift count is six bits with the REX.W prefix and five bits without.
1750 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
1751 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
1752 return Imm & ShiftCountMask;
1753}
1754
1755/// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
1756/// can be represented by a LEA instruction.
1757inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
1758 // Left shift instructions can be transformed into load-effective-address
1759 // instructions if we can encode them appropriately.
1760 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
1761 // The SIB.scale field is two bits wide which means that we can encode any
1762 // shift amount less than 4.
1763 return ShAmt < 4 && ShAmt > 0;
1764}
1765
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1766/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1767/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1768/// to a 32-bit superregister and then truncating back down to a 16-bit
1769/// subregister.
1770MachineInstr *
1771X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1772 MachineFunction::iterator &MFI,
1773 MachineBasicBlock::iterator &MBBI,
1774 LiveVariables *LV) const {
1775 MachineInstr *MI = MBBI;
1776 unsigned Dest = MI->getOperand(0).getReg();
1777 unsigned Src = MI->getOperand(1).getReg();
1778 bool isDead = MI->getOperand(0).isDead();
1779 bool isKill = MI->getOperand(1).isKill();
1780
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]1781 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1782 ? X86::LEA64_32r : X86::LEA32r;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1783 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]1784 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1785 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1786
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1787 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1788 // well be shifting and then extracting the lower 16-bits.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1789 // This has the potential to cause partial register stall. e.g.
Evan Cheng3974c8d2009-12-12 18:55:26 +0000[diff] [blame]1790 // movw (%rbp,%rcx,2), %dx
1791 // leal -65(%rdx), %esi
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1792 // But testing has shown this *does* help performance in 64-bit mode (at
1793 // least on modern x86 machines).
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1794 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1795 MachineInstr *InsMI =
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1796 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1797 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1798 .addReg(Src, getKillRegState(isKill));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1799
1800 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1801 get(Opc), leaOutReg);
1802 switch (MIOpc) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1803 default: llvm_unreachable("Unreachable!");
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1804 case X86::SHL16ri: {
1805 unsigned ShAmt = MI->getOperand(2).getImm();
1806 MIB.addReg(0).addImm(1 << ShAmt)
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1807 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1808 break;
1809 }
1810 case X86::INC16r:
1811 case X86::INC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1812 addRegOffset(MIB, leaInReg, true, 1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1813 break;
1814 case X86::DEC16r:
1815 case X86::DEC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1816 addRegOffset(MIB, leaInReg, true, -1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1817 break;
1818 case X86::ADD16ri:
1819 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]1820 case X86::ADD16ri_DB:
1821 case X86::ADD16ri8_DB:
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1822 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1823 break;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1824 case X86::ADD16rr:
1825 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1826 unsigned Src2 = MI->getOperand(2).getReg();
1827 bool isKill2 = MI->getOperand(2).isKill();
1828 unsigned leaInReg2 = 0;
1829 MachineInstr *InsMI2 = 0;
1830 if (Src == Src2) {
1831 // ADD16rr %reg1028<kill>, %reg1028
1832 // just a single insert_subreg.
1833 addRegReg(MIB, leaInReg, true, leaInReg, false);
1834 } else {
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]1835 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1836 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1837 // well be shifting and then extracting the lower 16-bits.
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1838 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1839 InsMI2 =
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1840 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1841 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1842 .addReg(Src2, getKillRegState(isKill2));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1843 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1844 }
1845 if (LV && isKill2 && InsMI2)
1846 LV->replaceKillInstruction(Src2, MI, InsMI2);
1847 break;
1848 }
1849 }
1850
1851 MachineInstr *NewMI = MIB;
1852 MachineInstr *ExtMI =
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1853 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1854 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1855 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1856
1857 if (LV) {
1858 // Update live variables
1859 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1860 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1861 if (isKill)
1862 LV->replaceKillInstruction(Src, MI, InsMI);
1863 if (isDead)
1864 LV->replaceKillInstruction(Dest, MI, ExtMI);
1865 }
1866
1867 return ExtMI;
1868}
1869
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1870/// convertToThreeAddress - This method must be implemented by targets that
1871/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1872/// may be able to convert a two-address instruction into a true
1873/// three-address instruction on demand. This allows the X86 target (for
1874/// example) to convert ADD and SHL instructions into LEA instructions if they
1875/// would require register copies due to two-addressness.
1876///
1877/// This method returns a null pointer if the transformation cannot be
1878/// performed, otherwise it returns the new instruction.
1879///
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1880MachineInstr *
1881X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1882 MachineBasicBlock::iterator &MBBI,
Owen Anderson30cc0282008-07-02 23:41:07 +0000[diff] [blame]1883 LiveVariables *LV) const {
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1884 MachineInstr *MI = MBBI;
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]1885
1886 // The following opcodes also sets the condition code register(s). Only
1887 // convert them to equivalent lea if the condition code register def's
1888 // are dead!
1889 if (hasLiveCondCodeDef(MI))
1890 return 0;
1891
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1892 MachineFunction &MF = *MI->getParent()->getParent();
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1893 // All instructions input are two-addr instructions. Get the known operands.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1894 const MachineOperand &Dest = MI->getOperand(0);
1895 const MachineOperand &Src = MI->getOperand(1);
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1896
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]1897 MachineInstr *NewMI = NULL;
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1898 // 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]1899 // we have better subtarget support, enable the 16-bit LEA generation here.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1900 // 16-bit LEA is also slow on Core2.
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1901 bool DisableLEA16 = true;
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1902 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1903
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1904 unsigned MIOpc = MI->getOpcode();
1905 switch (MIOpc) {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1906 case X86::SHUFPSrri: {
1907 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1908 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1909
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]1910 unsigned B = MI->getOperand(1).getReg();
1911 unsigned C = MI->getOperand(2).getReg();
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1912 if (B != C) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1913 unsigned M = MI->getOperand(3).getImm();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1914 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1915 .addOperand(Dest).addOperand(Src).addImm(M);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1916 break;
1917 }
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1918 case X86::SHUFPDrri: {
1919 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1920 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1921
1922 unsigned B = MI->getOperand(1).getReg();
1923 unsigned C = MI->getOperand(2).getReg();
1924 if (B != C) return 0;
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1925 unsigned M = MI->getOperand(3).getImm();
1926
1927 // Convert to PSHUFD mask.
1928 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1929
1930 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1931 .addOperand(Dest).addOperand(Src).addImm(M);
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1932 break;
1933 }
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1934 case X86::SHL64ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1935 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]1936 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1937 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1938
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1939 // LEA can't handle RSP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1940 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1941 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1942 &X86::GR64_NOSPRegClass))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1943 return 0;
1944
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1945 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1946 .addOperand(Dest)
1947 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1948 break;
1949 }
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1950 case X86::SHL32ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1951 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]1952 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1953 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1954
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1955 // LEA can't handle ESP.
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]1956 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1957 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1958 &X86::GR32_NOSPRegClass))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1959 return 0;
1960
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1961 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1962 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1963 .addOperand(Dest)
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]1964 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1965 break;
1966 }
1967 case X86::SHL16ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1968 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]1969 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1970 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1971
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1972 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1973 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1974 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1975 .addOperand(Dest)
1976 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1977 break;
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1978 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1979 default: {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1980
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1981 switch (MIOpc) {
1982 default: return 0;
1983 case X86::INC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]1984 case X86::INC32r:
1985 case X86::INC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1986 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]1987 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1988 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]1989 const TargetRegisterClass *RC = MIOpc == X86::INC64r ?
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]1990 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1991 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1992
1993 // LEA can't handle RSP.
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]1994 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1995 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1996 return 0;
1997
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1998 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]1999 .addOperand(Dest).addOperand(Src), 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2000 break;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2001 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2002 case X86::INC16r:
2003 case X86::INC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2004 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2005 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2006 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2007 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2008 .addOperand(Dest).addOperand(Src), 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2009 break;
2010 case X86::DEC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]2011 case X86::DEC32r:
2012 case X86::DEC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2013 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]2014 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2015 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2016 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ?
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]2017 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
2018 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2019 // LEA can't handle RSP.
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2020 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2021 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2022 return 0;
2023
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2024 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2025 .addOperand(Dest).addOperand(Src), -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2026 break;
2027 }
2028 case X86::DEC16r:
2029 case X86::DEC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2030 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2031 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2032 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2033 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2034 .addOperand(Dest).addOperand(Src), -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2035 break;
2036 case X86::ADD64rr:
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2037 case X86::ADD64rr_DB:
2038 case X86::ADD32rr:
2039 case X86::ADD32rr_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2040 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2041 unsigned Opc;
Craig Topper760b1342012-02-22 05:59:10 +0000[diff] [blame]2042 const TargetRegisterClass *RC;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2043 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
2044 Opc = X86::LEA64r;
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]2045 RC = &X86::GR64_NOSPRegClass;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2046 } else {
2047 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2048 RC = &X86::GR32_NOSPRegClass;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2049 }
2050
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2051
2052 unsigned Src2 = MI->getOperand(2).getReg();
2053 bool isKill2 = MI->getOperand(2).isKill();
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2054
2055 // LEA can't handle RSP.
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2056 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
2057 !MF.getRegInfo().constrainRegClass(Src2, RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2058 return 0;
2059
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2060 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2061 .addOperand(Dest),
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2062 Src.getReg(), Src.isKill(), Src2, isKill2);
Nadav Rotem4968e452012-07-16 10:52:25 +0000[diff] [blame]2063
2064 // Preserve undefness of the operands.
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2065 bool isUndef = MI->getOperand(1).isUndef();
2066 bool isUndef2 = MI->getOperand(2).isUndef();
2067 NewMI->getOperand(1).setIsUndef(isUndef);
2068 NewMI->getOperand(3).setIsUndef(isUndef2);
Nadav Rotem4968e452012-07-16 10:52:25 +0000[diff] [blame]2069
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2070 if (LV && isKill2)
2071 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2072 break;
2073 }
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2074 case X86::ADD16rr:
2075 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2076 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2077 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2078 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2079 unsigned Src2 = MI->getOperand(2).getReg();
2080 bool isKill2 = MI->getOperand(2).isKill();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2081 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2082 .addOperand(Dest),
2083 Src.getReg(), Src.isKill(), Src2, isKill2);
2084
2085 // Preserve undefness of the operands.
2086 bool isUndef = MI->getOperand(1).isUndef();
2087 bool isUndef2 = MI->getOperand(2).isUndef();
2088 NewMI->getOperand(1).setIsUndef(isUndef);
2089 NewMI->getOperand(3).setIsUndef(isUndef2);
2090
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2091 if (LV && isKill2)
2092 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2093 break;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2094 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2095 case X86::ADD64ri32:
2096 case X86::ADD64ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2097 case X86::ADD64ri32_DB:
2098 case X86::ADD64ri8_DB:
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2099 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2100 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2101 .addOperand(Dest).addOperand(Src),
2102 MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2103 break;
2104 case X86::ADD32ri:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2105 case X86::ADD32ri8:
2106 case X86::ADD32ri_DB:
2107 case X86::ADD32ri8_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2108 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Tim Northover339bf152013-06-01 10:23:46 +0000[diff] [blame]2109 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2110 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2111 .addOperand(Dest).addOperand(Src),
2112 MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2113 break;
2114 }
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2115 case X86::ADD16ri:
2116 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2117 case X86::ADD16ri_DB:
2118 case X86::ADD16ri8_DB:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2119 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2120 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2121 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2122 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2123 .addOperand(Dest).addOperand(Src),
2124 MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2125 break;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2126 }
2127 }
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2128 }
2129
Evan Cheng1bc1cae2008-02-07 08:29:53 +0000[diff] [blame]2130 if (!NewMI) return 0;
2131
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2132 if (LV) { // Update live variables
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2133 if (Src.isKill())
2134 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2135 if (Dest.isDead())
2136 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2137 }
2138
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2139 MFI->insert(MBBI, NewMI); // Insert the new inst
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]2140 return NewMI;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2141}
2142
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2143/// commuteInstruction - We have a few instructions that must be hacked on to
2144/// commute them.
2145///
Evan Cheng03553bb2008-06-16 07:33:11 +0000[diff] [blame]2146MachineInstr *
2147X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2148 switch (MI->getOpcode()) {
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2149 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2150 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2151 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2152 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2153 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2154 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2155 unsigned Opc;
2156 unsigned Size;
2157 switch (MI->getOpcode()) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2158 default: llvm_unreachable("Unreachable!");
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2159 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2160 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2161 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2162 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2163 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2164 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2165 }
Chris Lattner5c463782007-12-30 20:49:49 +0000[diff] [blame]2166 unsigned Amt = MI->getOperand(3).getImm();
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2167 if (NewMI) {
2168 MachineFunction &MF = *MI->getParent()->getParent();
2169 MI = MF.CloneMachineInstr(MI);
2170 NewMI = false;
Evan Cheng244183e2008-02-13 02:46:49 +0000[diff] [blame]2171 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2172 MI->setDesc(get(Opc));
2173 MI->getOperand(3).setImm(Size-Amt);
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]2174 return TargetInstrInfo::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2175 }
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2176 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2177 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2178 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2179 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2180 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2181 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2182 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2183 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2184 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2185 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2186 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2187 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2188 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2189 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2190 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2191 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2192 unsigned Opc;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2193 switch (MI->getOpcode()) {
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2194 default: llvm_unreachable("Unreachable!");
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2195 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2196 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2197 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2198 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2199 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2200 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2201 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2202 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2203 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2204 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2205 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2206 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
Chris Lattner1a1c6002010-10-05 23:00:14 +0000[diff] [blame]2207 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2208 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2209 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2210 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2211 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2212 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2213 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2214 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2215 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2216 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2217 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2218 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2219 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2220 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2221 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2222 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2223 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2224 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2225 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2226 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2227 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2228 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2229 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2230 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2231 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2232 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2233 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2234 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2235 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2236 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2237 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2238 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2239 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2240 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2241 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2242 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2243 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2244 if (NewMI) {
2245 MachineFunction &MF = *MI->getParent()->getParent();
2246 MI = MF.CloneMachineInstr(MI);
2247 NewMI = false;
2248 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]2249 MI->setDesc(get(Opc));
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2250 // Fallthrough intended.
2251 }
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2252 default:
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]2253 return TargetInstrInfo::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2254 }
2255}
2256
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2257static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2258 switch (BrOpc) {
2259 default: return X86::COND_INVALID;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2260 case X86::JE_4: return X86::COND_E;
2261 case X86::JNE_4: return X86::COND_NE;
2262 case X86::JL_4: return X86::COND_L;
2263 case X86::JLE_4: return X86::COND_LE;
2264 case X86::JG_4: return X86::COND_G;
2265 case X86::JGE_4: return X86::COND_GE;
2266 case X86::JB_4: return X86::COND_B;
2267 case X86::JBE_4: return X86::COND_BE;
2268 case X86::JA_4: return X86::COND_A;
2269 case X86::JAE_4: return X86::COND_AE;
2270 case X86::JS_4: return X86::COND_S;
2271 case X86::JNS_4: return X86::COND_NS;
2272 case X86::JP_4: return X86::COND_P;
2273 case X86::JNP_4: return X86::COND_NP;
2274 case X86::JO_4: return X86::COND_O;
2275 case X86::JNO_4: return X86::COND_NO;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2276 }
2277}
2278
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2279/// getCondFromSETOpc - return condition code of a SET opcode.
2280static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2281 switch (Opc) {
2282 default: return X86::COND_INVALID;
2283 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2284 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2285 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2286 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2287 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2288 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2289 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2290 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2291 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2292 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2293 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2294 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2295 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2296 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2297 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2298 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2299 }
2300}
2301
2302/// getCondFromCmovOpc - return condition code of a CMov opcode.
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]2303X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2304 switch (Opc) {
2305 default: return X86::COND_INVALID;
2306 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2307 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2308 return X86::COND_A;
2309 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2310 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2311 return X86::COND_AE;
2312 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2313 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2314 return X86::COND_B;
2315 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2316 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2317 return X86::COND_BE;
2318 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2319 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2320 return X86::COND_E;
2321 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2322 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2323 return X86::COND_G;
2324 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2325 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2326 return X86::COND_GE;
2327 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2328 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2329 return X86::COND_L;
2330 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2331 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2332 return X86::COND_LE;
2333 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2334 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2335 return X86::COND_NE;
2336 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2337 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2338 return X86::COND_NO;
2339 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2340 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2341 return X86::COND_NP;
2342 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2343 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2344 return X86::COND_NS;
2345 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2346 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2347 return X86::COND_O;
2348 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2349 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2350 return X86::COND_P;
2351 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2352 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2353 return X86::COND_S;
2354 }
2355}
2356
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2357unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2358 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2359 default: llvm_unreachable("Illegal condition code!");
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2360 case X86::COND_E: return X86::JE_4;
2361 case X86::COND_NE: return X86::JNE_4;
2362 case X86::COND_L: return X86::JL_4;
2363 case X86::COND_LE: return X86::JLE_4;
2364 case X86::COND_G: return X86::JG_4;
2365 case X86::COND_GE: return X86::JGE_4;
2366 case X86::COND_B: return X86::JB_4;
2367 case X86::COND_BE: return X86::JBE_4;
2368 case X86::COND_A: return X86::JA_4;
2369 case X86::COND_AE: return X86::JAE_4;
2370 case X86::COND_S: return X86::JS_4;
2371 case X86::COND_NS: return X86::JNS_4;
2372 case X86::COND_P: return X86::JP_4;
2373 case X86::COND_NP: return X86::JNP_4;
2374 case X86::COND_O: return X86::JO_4;
2375 case X86::COND_NO: return X86::JNO_4;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2376 }
2377}
2378
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2379/// GetOppositeBranchCondition - Return the inverse of the specified condition,
2380/// e.g. turning COND_E to COND_NE.
2381X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2382 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2383 default: llvm_unreachable("Illegal condition code!");
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2384 case X86::COND_E: return X86::COND_NE;
2385 case X86::COND_NE: return X86::COND_E;
2386 case X86::COND_L: return X86::COND_GE;
2387 case X86::COND_LE: return X86::COND_G;
2388 case X86::COND_G: return X86::COND_LE;
2389 case X86::COND_GE: return X86::COND_L;
2390 case X86::COND_B: return X86::COND_AE;
2391 case X86::COND_BE: return X86::COND_A;
2392 case X86::COND_A: return X86::COND_BE;
2393 case X86::COND_AE: return X86::COND_B;
2394 case X86::COND_S: return X86::COND_NS;
2395 case X86::COND_NS: return X86::COND_S;
2396 case X86::COND_P: return X86::COND_NP;
2397 case X86::COND_NP: return X86::COND_P;
2398 case X86::COND_O: return X86::COND_NO;
2399 case X86::COND_NO: return X86::COND_O;
2400 }
2401}
2402
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2403/// getSwappedCondition - assume the flags are set by MI(a,b), return
2404/// the condition code if we modify the instructions such that flags are
2405/// set by MI(b,a).
Benjamin Kramerabbfe692012-07-13 13:25:15 +0000[diff] [blame]2406static X86::CondCode getSwappedCondition(X86::CondCode CC) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2407 switch (CC) {
2408 default: return X86::COND_INVALID;
2409 case X86::COND_E: return X86::COND_E;
2410 case X86::COND_NE: return X86::COND_NE;
2411 case X86::COND_L: return X86::COND_G;
2412 case X86::COND_LE: return X86::COND_GE;
2413 case X86::COND_G: return X86::COND_L;
2414 case X86::COND_GE: return X86::COND_LE;
2415 case X86::COND_B: return X86::COND_A;
2416 case X86::COND_BE: return X86::COND_AE;
2417 case X86::COND_A: return X86::COND_B;
2418 case X86::COND_AE: return X86::COND_BE;
2419 }
2420}
2421
2422/// getSETFromCond - Return a set opcode for the given condition and
2423/// whether it has memory operand.
2424static unsigned getSETFromCond(X86::CondCode CC,
2425 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2426 static const uint16_t Opc[16][2] = {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2427 { X86::SETAr, X86::SETAm },
2428 { X86::SETAEr, X86::SETAEm },
2429 { X86::SETBr, X86::SETBm },
2430 { X86::SETBEr, X86::SETBEm },
2431 { X86::SETEr, X86::SETEm },
2432 { X86::SETGr, X86::SETGm },
2433 { X86::SETGEr, X86::SETGEm },
2434 { X86::SETLr, X86::SETLm },
2435 { X86::SETLEr, X86::SETLEm },
2436 { X86::SETNEr, X86::SETNEm },
2437 { X86::SETNOr, X86::SETNOm },
2438 { X86::SETNPr, X86::SETNPm },
2439 { X86::SETNSr, X86::SETNSm },
2440 { X86::SETOr, X86::SETOm },
2441 { X86::SETPr, X86::SETPm },
2442 { X86::SETSr, X86::SETSm }
2443 };
2444
2445 assert(CC < 16 && "Can only handle standard cond codes");
2446 return Opc[CC][HasMemoryOperand ? 1 : 0];
2447}
2448
2449/// getCMovFromCond - Return a cmov opcode for the given condition,
2450/// register size in bytes, and operand type.
2451static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2452 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2453 static const uint16_t Opc[32][3] = {
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2454 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2455 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2456 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2457 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2458 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2459 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2460 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2461 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2462 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2463 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2464 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2465 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2466 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2467 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2468 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2469 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2470 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2471 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2472 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2473 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2474 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2475 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2476 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2477 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2478 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2479 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2480 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2481 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2482 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2483 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2484 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2485 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2486 };
2487
2488 assert(CC < 16 && "Can only handle standard cond codes");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2489 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2490 switch(RegBytes) {
2491 default: llvm_unreachable("Illegal register size!");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2492 case 2: return Opc[Idx][0];
2493 case 4: return Opc[Idx][1];
2494 case 8: return Opc[Idx][2];
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2495 }
2496}
2497
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2498bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2499 if (!MI->isTerminator()) return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2500
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2501 // Conditional branch is a special case.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2502 if (MI->isBranch() && !MI->isBarrier())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2503 return true;
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2504 if (!MI->isPredicable())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2505 return true;
2506 return !isPredicated(MI);
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2507}
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2508
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2509bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2510 MachineBasicBlock *&TBB,
2511 MachineBasicBlock *&FBB,
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2512 SmallVectorImpl<MachineOperand> &Cond,
2513 bool AllowModify) const {
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2514 // Start from the bottom of the block and work up, examining the
2515 // terminator instructions.
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2516 MachineBasicBlock::iterator I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2517 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2518 while (I != MBB.begin()) {
2519 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2520 if (I->isDebugValue())
2521 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2522
2523 // Working from the bottom, when we see a non-terminator instruction, we're
2524 // done.
Jakob Stoklund Olesenc30b4dd2010-07-16 17:41:44 +0000[diff] [blame]2525 if (!isUnpredicatedTerminator(I))
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2526 break;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2527
2528 // A terminator that isn't a branch can't easily be handled by this
2529 // analysis.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2530 if (!I->isBranch())
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2531 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2532
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2533 // Handle unconditional branches.
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2534 if (I->getOpcode() == X86::JMP_4) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2535 UnCondBrIter = I;
2536
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2537 if (!AllowModify) {
2538 TBB = I->getOperand(0).getMBB();
Evan Cheng2fa28112009-05-08 06:34:09 +0000[diff] [blame]2539 continue;
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2540 }
2541
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2542 // If the block has any instructions after a JMP, delete them.
Chris Lattnera48f44d2009-12-03 00:50:42 +0000[diff] [blame]2543 while (llvm::next(I) != MBB.end())
2544 llvm::next(I)->eraseFromParent();
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2545
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2546 Cond.clear();
2547 FBB = 0;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2548
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2549 // Delete the JMP if it's equivalent to a fall-through.
2550 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2551 TBB = 0;
2552 I->eraseFromParent();
2553 I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2554 UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2555 continue;
2556 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2557
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2558 // TBB is used to indicate the unconditional destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2559 TBB = I->getOperand(0).getMBB();
2560 continue;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2561 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2562
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2563 // Handle conditional branches.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2564 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2565 if (BranchCode == X86::COND_INVALID)
2566 return true; // Can't handle indirect branch.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2567
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2568 // Working from the bottom, handle the first conditional branch.
2569 if (Cond.empty()) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2570 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2571 if (AllowModify && UnCondBrIter != MBB.end() &&
2572 MBB.isLayoutSuccessor(TargetBB)) {
2573 // If we can modify the code and it ends in something like:
2574 //
2575 // jCC L1
2576 // jmp L2
2577 // L1:
2578 // ...
2579 // L2:
2580 //
2581 // Then we can change this to:
2582 //
2583 // jnCC L2
2584 // L1:
2585 // ...
2586 // L2:
2587 //
2588 // Which is a bit more efficient.
2589 // We conditionally jump to the fall-through block.
2590 BranchCode = GetOppositeBranchCondition(BranchCode);
2591 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2592 MachineBasicBlock::iterator OldInst = I;
2593
2594 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2595 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2596 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2597 .addMBB(TargetBB);
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2598
2599 OldInst->eraseFromParent();
2600 UnCondBrIter->eraseFromParent();
2601
2602 // Restart the analysis.
2603 UnCondBrIter = MBB.end();
2604 I = MBB.end();
2605 continue;
2606 }
2607
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2608 FBB = TBB;
2609 TBB = I->getOperand(0).getMBB();
2610 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2611 continue;
2612 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2613
2614 // Handle subsequent conditional branches. Only handle the case where all
2615 // conditional branches branch to the same destination and their condition
2616 // opcodes fit one of the special multi-branch idioms.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2617 assert(Cond.size() == 1);
2618 assert(TBB);
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2619
2620 // Only handle the case where all conditional branches branch to the same
2621 // destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2622 if (TBB != I->getOperand(0).getMBB())
2623 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2624
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2625 // If the conditions are the same, we can leave them alone.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2626 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2627 if (OldBranchCode == BranchCode)
2628 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2629
2630 // If they differ, see if they fit one of the known patterns. Theoretically,
2631 // we could handle more patterns here, but we shouldn't expect to see them
2632 // if instruction selection has done a reasonable job.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2633 if ((OldBranchCode == X86::COND_NP &&
2634 BranchCode == X86::COND_E) ||
2635 (OldBranchCode == X86::COND_E &&
2636 BranchCode == X86::COND_NP))
2637 BranchCode = X86::COND_NP_OR_E;
2638 else if ((OldBranchCode == X86::COND_P &&
2639 BranchCode == X86::COND_NE) ||
2640 (OldBranchCode == X86::COND_NE &&
2641 BranchCode == X86::COND_P))
2642 BranchCode = X86::COND_NE_OR_P;
2643 else
2644 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2645
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2646 // Update the MachineOperand.
2647 Cond[0].setImm(BranchCode);
Chris Lattner74436002006-10-30 22:27:23 +0000[diff] [blame]2648 }
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2649
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2650 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2651}
2652
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2653unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2654 MachineBasicBlock::iterator I = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2655 unsigned Count = 0;
2656
2657 while (I != MBB.begin()) {
2658 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2659 if (I->isDebugValue())
2660 continue;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2661 if (I->getOpcode() != X86::JMP_4 &&
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2662 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2663 break;
2664 // Remove the branch.
2665 I->eraseFromParent();
2666 I = MBB.end();
2667 ++Count;
2668 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2669
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2670 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2671}
2672
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2673unsigned
2674X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2675 MachineBasicBlock *FBB,
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2676 const SmallVectorImpl<MachineOperand> &Cond,
2677 DebugLoc DL) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2678 // Shouldn't be a fall through.
2679 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
Chris Lattner6fca75e2006-10-21 05:34:23 +0000[diff] [blame]2680 assert((Cond.size() == 1 || Cond.size() == 0) &&
2681 "X86 branch conditions have one component!");
2682
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2683 if (Cond.empty()) {
2684 // Unconditional branch?
2685 assert(!FBB && "Unconditional branch with multiple successors!");
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2686 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2687 return 1;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2688 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2689
2690 // Conditional branch.
2691 unsigned Count = 0;
2692 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2693 switch (CC) {
2694 case X86::COND_NP_OR_E:
2695 // Synthesize NP_OR_E with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2696 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2697 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2698 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2699 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2700 break;
2701 case X86::COND_NE_OR_P:
2702 // Synthesize NE_OR_P with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2703 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2704 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2705 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2706 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2707 break;
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2708 default: {
2709 unsigned Opc = GetCondBranchFromCond(CC);
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2710 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2711 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2712 }
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2713 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2714 if (FBB) {
2715 // Two-way Conditional branch. Insert the second branch.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2716 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2717 ++Count;
2718 }
2719 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2720}
2721
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2722bool X86InstrInfo::
2723canInsertSelect(const MachineBasicBlock &MBB,
2724 const SmallVectorImpl<MachineOperand> &Cond,
2725 unsigned TrueReg, unsigned FalseReg,
2726 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2727 // Not all subtargets have cmov instructions.
2728 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2729 return false;
2730 if (Cond.size() != 1)
2731 return false;
2732 // We cannot do the composite conditions, at least not in SSA form.
2733 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2734 return false;
2735
2736 // Check register classes.
2737 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2738 const TargetRegisterClass *RC =
2739 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2740 if (!RC)
2741 return false;
2742
2743 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2744 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2745 X86::GR32RegClass.hasSubClassEq(RC) ||
2746 X86::GR64RegClass.hasSubClassEq(RC)) {
2747 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2748 // Bridge. Probably Ivy Bridge as well.
2749 CondCycles = 2;
2750 TrueCycles = 2;
2751 FalseCycles = 2;
2752 return true;
2753 }
2754
2755 // Can't do vectors.
2756 return false;
2757}
2758
2759void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2760 MachineBasicBlock::iterator I, DebugLoc DL,
2761 unsigned DstReg,
2762 const SmallVectorImpl<MachineOperand> &Cond,
2763 unsigned TrueReg, unsigned FalseReg) const {
2764 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2765 assert(Cond.size() == 1 && "Invalid Cond array");
2766 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2767 MRI.getRegClass(DstReg)->getSize(),
2768 false/*HasMemoryOperand*/);
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2769 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2770}
2771
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2772/// isHReg - Test if the given register is a physical h register.
2773static bool isHReg(unsigned Reg) {
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2774 return X86::GR8_ABCD_HRegClass.contains(Reg);
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2775}
2776
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2777// Try and copy between VR128/VR64 and GR64 registers.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2778static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2779 bool HasAVX) {
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2780 // SrcReg(VR128) -> DestReg(GR64)
2781 // SrcReg(VR64) -> DestReg(GR64)
2782 // SrcReg(GR64) -> DestReg(VR128)
2783 // SrcReg(GR64) -> DestReg(VR64)
2784
2785 if (X86::GR64RegClass.contains(DestReg)) {
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2786 if (X86::VR128RegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2787 // Copy from a VR128 register to a GR64 register.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2788 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2789 if (X86::VR64RegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2790 // Copy from a VR64 register to a GR64 register.
2791 return X86::MOVSDto64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2792 } else if (X86::GR64RegClass.contains(SrcReg)) {
2793 // Copy from a GR64 register to a VR128 register.
2794 if (X86::VR128RegClass.contains(DestReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2795 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2796 // Copy from a GR64 register to a VR64 register.
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2797 if (X86::VR64RegClass.contains(DestReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2798 return X86::MOV64toSDrr;
2799 }
2800
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2801 // SrcReg(FR32) -> DestReg(GR32)
2802 // SrcReg(GR32) -> DestReg(FR32)
2803
2804 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2805 // Copy from a FR32 register to a GR32 register.
2806 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2807
2808 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2809 // Copy from a GR32 register to a FR32 register.
2810 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2811
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2812 return 0;
2813}
2814
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2815void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2816 MachineBasicBlock::iterator MI, DebugLoc DL,
2817 unsigned DestReg, unsigned SrcReg,
2818 bool KillSrc) const {
2819 // First deal with the normal symmetric copies.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2820 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2821 unsigned Opc;
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2822 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2823 Opc = X86::MOV64rr;
2824 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2825 Opc = X86::MOV32rr;
2826 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2827 Opc = X86::MOV16rr;
2828 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2829 // Copying to or from a physical H register on x86-64 requires a NOREX
2830 // move. Otherwise use a normal move.
2831 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2832 TM.getSubtarget<X86Subtarget>().is64Bit()) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2833 Opc = X86::MOV8rr_NOREX;
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2834 // Both operands must be encodable without an REX prefix.
2835 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2836 "8-bit H register can not be copied outside GR8_NOREX");
2837 } else
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2838 Opc = X86::MOV8rr;
2839 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2840 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]2841 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2842 Opc = X86::VMOVAPSYrr;
Jakob Stoklund Olesenec58a432010-07-08 22:30:35 +0000[diff] [blame]2843 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2844 Opc = X86::MMX_MOVQ64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2845 else
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2846 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2847
2848 if (Opc) {
2849 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2850 .addReg(SrcReg, getKillRegState(KillSrc));
2851 return;
2852 }
2853
2854 // Moving EFLAGS to / from another register requires a push and a pop.
Nadav Rotemd5aae982012-12-21 23:48:49 +0000[diff] [blame]2855 // Notice that we have to adjust the stack if we don't want to clobber the
2856 // first frame index. See X86FrameLowering.cpp - colobbersTheStack.
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2857 if (SrcReg == X86::EFLAGS) {
2858 if (X86::GR64RegClass.contains(DestReg)) {
2859 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2860 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2861 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2862 }
2863 if (X86::GR32RegClass.contains(DestReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2864 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2865 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2866 return;
2867 }
2868 }
2869 if (DestReg == X86::EFLAGS) {
2870 if (X86::GR64RegClass.contains(SrcReg)) {
2871 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2872 .addReg(SrcReg, getKillRegState(KillSrc));
2873 BuildMI(MBB, MI, DL, get(X86::POPF64));
2874 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2875 }
2876 if (X86::GR32RegClass.contains(SrcReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2877 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2878 .addReg(SrcReg, getKillRegState(KillSrc));
2879 BuildMI(MBB, MI, DL, get(X86::POPF32));
2880 return;
2881 }
2882 }
2883
2884 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2885 << " to " << RI.getName(DestReg) << '\n');
2886 llvm_unreachable("Cannot emit physreg copy instruction");
2887}
2888
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2889static unsigned getLoadStoreRegOpcode(unsigned Reg,
2890 const TargetRegisterClass *RC,
2891 bool isStackAligned,
2892 const TargetMachine &TM,
2893 bool load) {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2894 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2895 switch (RC->getSize()) {
Rafael Espindola6635f982010-07-12 03:43:04 +0000[diff] [blame]2896 default:
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2897 llvm_unreachable("Unknown spill size");
2898 case 1:
2899 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2900 if (TM.getSubtarget<X86Subtarget>().is64Bit())
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2901 // Copying to or from a physical H register on x86-64 requires a NOREX
2902 // move. Otherwise use a normal move.
2903 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2904 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2905 return load ? X86::MOV8rm : X86::MOV8mr;
2906 case 2:
2907 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2908 return load ? X86::MOV16rm : X86::MOV16mr;
2909 case 4:
2910 if (X86::GR32RegClass.hasSubClassEq(RC))
2911 return load ? X86::MOV32rm : X86::MOV32mr;
2912 if (X86::FR32RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2913 return load ?
2914 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2915 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2916 if (X86::RFP32RegClass.hasSubClassEq(RC))
2917 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2918 llvm_unreachable("Unknown 4-byte regclass");
2919 case 8:
2920 if (X86::GR64RegClass.hasSubClassEq(RC))
2921 return load ? X86::MOV64rm : X86::MOV64mr;
2922 if (X86::FR64RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2923 return load ?
2924 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
2925 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2926 if (X86::VR64RegClass.hasSubClassEq(RC))
2927 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2928 if (X86::RFP64RegClass.hasSubClassEq(RC))
2929 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2930 llvm_unreachable("Unknown 8-byte regclass");
2931 case 10:
2932 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2933 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2934 case 16: {
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2935 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2936 // If stack is realigned we can use aligned stores.
2937 if (isStackAligned)
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2938 return load ?
2939 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
2940 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2941 else
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2942 return load ?
2943 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
2944 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
2945 }
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]2946 case 32:
2947 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2948 // If stack is realigned we can use aligned stores.
2949 if (isStackAligned)
2950 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2951 else
2952 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2953 }
2954}
2955
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2956static unsigned getStoreRegOpcode(unsigned SrcReg,
2957 const TargetRegisterClass *RC,
2958 bool isStackAligned,
2959 TargetMachine &TM) {
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2960 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2961}
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2962
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2963
2964static unsigned getLoadRegOpcode(unsigned DestReg,
2965 const TargetRegisterClass *RC,
2966 bool isStackAligned,
2967 const TargetMachine &TM) {
2968 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2969}
2970
2971void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
2972 MachineBasicBlock::iterator MI,
2973 unsigned SrcReg, bool isKill, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]2974 const TargetRegisterClass *RC,
2975 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]2976 const MachineFunction &MF = *MBB.getParent();
Jakob Stoklund Olesenc3c05ed2010-07-27 04:16:58 +0000[diff] [blame]2977 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
2978 "Stack slot too small for store");
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2979 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2980 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]2981 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2982 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]2983 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2984 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]2985 .addReg(SrcReg, getKillRegState(isKill));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2986}
2987
2988void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
2989 bool isKill,
2990 SmallVectorImpl<MachineOperand> &Addr,
2991 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]2992 MachineInstr::mmo_iterator MMOBegin,
2993 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2994 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2995 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2996 bool isAligned = MMOBegin != MMOEnd &&
2997 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2998 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]2999 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]3000 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3001 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3002 MIB.addOperand(Addr[i]);
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]3003 MIB.addReg(SrcReg, getKillRegState(isKill));
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3004 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3005 NewMIs.push_back(MIB);
3006}
3007
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3008
3009void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3010 MachineBasicBlock::iterator MI,
3011 unsigned DestReg, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]3012 const TargetRegisterClass *RC,
3013 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3014 const MachineFunction &MF = *MBB.getParent();
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3015 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3016 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]3017 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3018 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]3019 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3020 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3021}
3022
3023void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]3024 SmallVectorImpl<MachineOperand> &Addr,
3025 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3026 MachineInstr::mmo_iterator MMOBegin,
3027 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3028 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3029 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3030 bool isAligned = MMOBegin != MMOEnd &&
3031 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3032 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]3033 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]3034 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3035 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3036 MIB.addOperand(Addr[i]);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3037 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3038 NewMIs.push_back(MIB);
3039}
3040
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3041bool X86InstrInfo::
3042analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3043 int &CmpMask, int &CmpValue) const {
3044 switch (MI->getOpcode()) {
3045 default: break;
3046 case X86::CMP64ri32:
3047 case X86::CMP64ri8:
3048 case X86::CMP32ri:
3049 case X86::CMP32ri8:
3050 case X86::CMP16ri:
3051 case X86::CMP16ri8:
3052 case X86::CMP8ri:
3053 SrcReg = MI->getOperand(0).getReg();
3054 SrcReg2 = 0;
3055 CmpMask = ~0;
3056 CmpValue = MI->getOperand(1).getImm();
3057 return true;
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3058 // A SUB can be used to perform comparison.
3059 case X86::SUB64rm:
3060 case X86::SUB32rm:
3061 case X86::SUB16rm:
3062 case X86::SUB8rm:
3063 SrcReg = MI->getOperand(1).getReg();
3064 SrcReg2 = 0;
3065 CmpMask = ~0;
3066 CmpValue = 0;
3067 return true;
3068 case X86::SUB64rr:
3069 case X86::SUB32rr:
3070 case X86::SUB16rr:
3071 case X86::SUB8rr:
3072 SrcReg = MI->getOperand(1).getReg();
3073 SrcReg2 = MI->getOperand(2).getReg();
3074 CmpMask = ~0;
3075 CmpValue = 0;
3076 return true;
3077 case X86::SUB64ri32:
3078 case X86::SUB64ri8:
3079 case X86::SUB32ri:
3080 case X86::SUB32ri8:
3081 case X86::SUB16ri:
3082 case X86::SUB16ri8:
3083 case X86::SUB8ri:
3084 SrcReg = MI->getOperand(1).getReg();
3085 SrcReg2 = 0;
3086 CmpMask = ~0;
3087 CmpValue = MI->getOperand(2).getImm();
3088 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3089 case X86::CMP64rr:
3090 case X86::CMP32rr:
3091 case X86::CMP16rr:
3092 case X86::CMP8rr:
3093 SrcReg = MI->getOperand(0).getReg();
3094 SrcReg2 = MI->getOperand(1).getReg();
3095 CmpMask = ~0;
3096 CmpValue = 0;
3097 return true;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3098 case X86::TEST8rr:
3099 case X86::TEST16rr:
3100 case X86::TEST32rr:
3101 case X86::TEST64rr:
3102 SrcReg = MI->getOperand(0).getReg();
3103 if (MI->getOperand(1).getReg() != SrcReg) return false;
3104 // Compare against zero.
3105 SrcReg2 = 0;
3106 CmpMask = ~0;
3107 CmpValue = 0;
3108 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3109 }
3110 return false;
3111}
3112
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3113/// isRedundantFlagInstr - check whether the first instruction, whose only
3114/// purpose is to update flags, can be made redundant.
3115/// CMPrr can be made redundant by SUBrr if the operands are the same.
3116/// This function can be extended later on.
3117/// SrcReg, SrcRegs: register operands for FlagI.
3118/// ImmValue: immediate for FlagI if it takes an immediate.
3119inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3120 unsigned SrcReg2, int ImmValue,
3121 MachineInstr *OI) {
3122 if (((FlagI->getOpcode() == X86::CMP64rr &&
3123 OI->getOpcode() == X86::SUB64rr) ||
3124 (FlagI->getOpcode() == X86::CMP32rr &&
3125 OI->getOpcode() == X86::SUB32rr)||
3126 (FlagI->getOpcode() == X86::CMP16rr &&
3127 OI->getOpcode() == X86::SUB16rr)||
3128 (FlagI->getOpcode() == X86::CMP8rr &&
3129 OI->getOpcode() == X86::SUB8rr)) &&
3130 ((OI->getOperand(1).getReg() == SrcReg &&
3131 OI->getOperand(2).getReg() == SrcReg2) ||
3132 (OI->getOperand(1).getReg() == SrcReg2 &&
3133 OI->getOperand(2).getReg() == SrcReg)))
3134 return true;
3135
3136 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3137 OI->getOpcode() == X86::SUB64ri32) ||
3138 (FlagI->getOpcode() == X86::CMP64ri8 &&
3139 OI->getOpcode() == X86::SUB64ri8) ||
3140 (FlagI->getOpcode() == X86::CMP32ri &&
3141 OI->getOpcode() == X86::SUB32ri) ||
3142 (FlagI->getOpcode() == X86::CMP32ri8 &&
3143 OI->getOpcode() == X86::SUB32ri8) ||
3144 (FlagI->getOpcode() == X86::CMP16ri &&
3145 OI->getOpcode() == X86::SUB16ri) ||
3146 (FlagI->getOpcode() == X86::CMP16ri8 &&
3147 OI->getOpcode() == X86::SUB16ri8) ||
3148 (FlagI->getOpcode() == X86::CMP8ri &&
3149 OI->getOpcode() == X86::SUB8ri)) &&
3150 OI->getOperand(1).getReg() == SrcReg &&
3151 OI->getOperand(2).getImm() == ImmValue)
3152 return true;
3153 return false;
3154}
3155
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3156/// isDefConvertible - check whether the definition can be converted
3157/// to remove a comparison against zero.
3158inline static bool isDefConvertible(MachineInstr *MI) {
3159 switch (MI->getOpcode()) {
3160 default: return false;
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame]3161
3162 // The shift instructions only modify ZF if their shift count is non-zero.
3163 // N.B.: The processor truncates the shift count depending on the encoding.
3164 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3165 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3166 return getTruncatedShiftCount(MI, 2) != 0;
3167
3168 // Some left shift instructions can be turned into LEA instructions but only
3169 // if their flags aren't used. Avoid transforming such instructions.
3170 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3171 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3172 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3173 return ShAmt != 0;
3174 }
3175
3176 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3177 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3178 return getTruncatedShiftCount(MI, 3) != 0;
3179
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3180 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3181 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3182 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3183 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3184 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
Craig Topper5b08cf72012-12-17 04:55:07 +0000[diff] [blame]3185 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3186 case X86::DEC64_32r: case X86::DEC64_16r:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3187 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3188 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3189 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3190 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3191 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
Craig Topper5b08cf72012-12-17 04:55:07 +0000[diff] [blame]3192 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3193 case X86::INC64_32r: case X86::INC64_16r:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3194 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3195 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3196 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3197 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3198 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3199 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3200 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3201 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3202 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3203 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3204 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3205 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3206 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3207 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3208 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3209 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3210 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3211 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3212 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3213 case X86::ADC32ri: case X86::ADC32ri8:
3214 case X86::ADC32rr: case X86::ADC64ri32:
3215 case X86::ADC64ri8: case X86::ADC64rr:
3216 case X86::SBB32ri: case X86::SBB32ri8:
3217 case X86::SBB32rr: case X86::SBB64ri32:
3218 case X86::SBB64ri8: case X86::SBB64rr:
Craig Topperf3ff6ae2012-12-17 05:12:30 +0000[diff] [blame]3219 case X86::ANDN32rr: case X86::ANDN32rm:
3220 case X86::ANDN64rr: case X86::ANDN64rm:
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3221 case X86::BEXTR32rr: case X86::BEXTR64rr:
3222 case X86::BEXTR32rm: case X86::BEXTR64rm:
3223 case X86::BLSI32rr: case X86::BLSI32rm:
3224 case X86::BLSI64rr: case X86::BLSI64rm:
3225 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3226 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3227 case X86::BLSR32rr: case X86::BLSR32rm:
3228 case X86::BLSR64rr: case X86::BLSR64rm:
3229 case X86::BZHI32rr: case X86::BZHI32rm:
3230 case X86::BZHI64rr: case X86::BZHI64rm:
3231 case X86::LZCNT16rr: case X86::LZCNT16rm:
3232 case X86::LZCNT32rr: case X86::LZCNT32rm:
3233 case X86::LZCNT64rr: case X86::LZCNT64rm:
3234 case X86::POPCNT16rr:case X86::POPCNT16rm:
3235 case X86::POPCNT32rr:case X86::POPCNT32rm:
3236 case X86::POPCNT64rr:case X86::POPCNT64rm:
3237 case X86::TZCNT16rr: case X86::TZCNT16rm:
3238 case X86::TZCNT32rr: case X86::TZCNT32rm:
3239 case X86::TZCNT64rr: case X86::TZCNT64rm:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3240 return true;
3241 }
3242}
3243
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3244/// optimizeCompareInstr - Check if there exists an earlier instruction that
3245/// operates on the same source operands and sets flags in the same way as
3246/// Compare; remove Compare if possible.
3247bool X86InstrInfo::
3248optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3249 int CmpMask, int CmpValue,
3250 const MachineRegisterInfo *MRI) const {
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3251 // Check whether we can replace SUB with CMP.
3252 unsigned NewOpcode = 0;
3253 switch (CmpInstr->getOpcode()) {
3254 default: break;
3255 case X86::SUB64ri32:
3256 case X86::SUB64ri8:
3257 case X86::SUB32ri:
3258 case X86::SUB32ri8:
3259 case X86::SUB16ri:
3260 case X86::SUB16ri8:
3261 case X86::SUB8ri:
3262 case X86::SUB64rm:
3263 case X86::SUB32rm:
3264 case X86::SUB16rm:
3265 case X86::SUB8rm:
3266 case X86::SUB64rr:
3267 case X86::SUB32rr:
3268 case X86::SUB16rr:
3269 case X86::SUB8rr: {
3270 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3271 return false;
3272 // There is no use of the destination register, we can replace SUB with CMP.
3273 switch (CmpInstr->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]3274 default: llvm_unreachable("Unreachable!");
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3275 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3276 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3277 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3278 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3279 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3280 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3281 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3282 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3283 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3284 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3285 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3286 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3287 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3288 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3289 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3290 }
3291 CmpInstr->setDesc(get(NewOpcode));
3292 CmpInstr->RemoveOperand(0);
3293 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3294 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3295 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3296 return false;
3297 }
3298 }
3299
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3300 // Get the unique definition of SrcReg.
3301 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3302 if (!MI) return false;
3303
3304 // CmpInstr is the first instruction of the BB.
3305 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3306
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3307 // If we are comparing against zero, check whether we can use MI to update
3308 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3309 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3310 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3311 !isDefConvertible(MI)))
3312 return false;
3313
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3314 // We are searching for an earlier instruction that can make CmpInstr
3315 // redundant and that instruction will be saved in Sub.
3316 MachineInstr *Sub = NULL;
3317 const TargetRegisterInfo *TRI = &getRegisterInfo();
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3318
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3319 // We iterate backward, starting from the instruction before CmpInstr and
3320 // stop when reaching the definition of a source register or done with the BB.
3321 // RI points to the instruction before CmpInstr.
3322 // If the definition is in this basic block, RE points to the definition;
3323 // otherwise, RE is the rend of the basic block.
3324 MachineBasicBlock::reverse_iterator
3325 RI = MachineBasicBlock::reverse_iterator(I),
3326 RE = CmpInstr->getParent() == MI->getParent() ?
3327 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3328 CmpInstr->getParent()->rend();
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3329 MachineInstr *Movr0Inst = 0;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3330 for (; RI != RE; ++RI) {
3331 MachineInstr *Instr = &*RI;
3332 // Check whether CmpInstr can be made redundant by the current instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3333 if (!IsCmpZero &&
3334 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3335 Sub = Instr;
3336 break;
3337 }
3338
3339 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3340 Instr->readsRegister(X86::EFLAGS, TRI)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3341 // This instruction modifies or uses EFLAGS.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3342
3343 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3344 // They are safe to move up, if the definition to EFLAGS is dead and
3345 // earlier instructions do not read or write EFLAGS.
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]3346 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3347 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3348 Movr0Inst = Instr;
3349 continue;
3350 }
3351
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3352 // We can't remove CmpInstr.
3353 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3354 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3355 }
3356
3357 // Return false if no candidates exist.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3358 if (!IsCmpZero && !Sub)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3359 return false;
3360
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3361 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3362 Sub->getOperand(2).getReg() == SrcReg);
3363
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3364 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3365 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3366 // If we are done with the basic block, we need to check whether EFLAGS is
3367 // live-out.
3368 bool IsSafe = false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3369 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3370 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3371 for (++I; I != E; ++I) {
3372 const MachineInstr &Instr = *I;
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3373 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3374 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3375 // We should check the usage if this instruction uses and updates EFLAGS.
3376 if (!UseEFLAGS && ModifyEFLAGS) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3377 // It is safe to remove CmpInstr if EFLAGS is updated again.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3378 IsSafe = true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3379 break;
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3380 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3381 if (!UseEFLAGS && !ModifyEFLAGS)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3382 continue;
3383
3384 // EFLAGS is used by this instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3385 X86::CondCode OldCC;
3386 bool OpcIsSET = false;
3387 if (IsCmpZero || IsSwapped) {
3388 // We decode the condition code from opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3389 if (Instr.isBranch())
3390 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3391 else {
3392 OldCC = getCondFromSETOpc(Instr.getOpcode());
3393 if (OldCC != X86::COND_INVALID)
3394 OpcIsSET = true;
3395 else
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]3396 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3397 }
3398 if (OldCC == X86::COND_INVALID) return false;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3399 }
3400 if (IsCmpZero) {
3401 switch (OldCC) {
3402 default: break;
3403 case X86::COND_A: case X86::COND_AE:
3404 case X86::COND_B: case X86::COND_BE:
3405 case X86::COND_G: case X86::COND_GE:
3406 case X86::COND_L: case X86::COND_LE:
3407 case X86::COND_O: case X86::COND_NO:
3408 // CF and OF are used, we can't perform this optimization.
3409 return false;
3410 }
3411 } else if (IsSwapped) {
3412 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3413 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3414 // We swap the condition code and synthesize the new opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3415 X86::CondCode NewCC = getSwappedCondition(OldCC);
3416 if (NewCC == X86::COND_INVALID) return false;
3417
3418 // Synthesize the new opcode.
3419 bool HasMemoryOperand = Instr.hasOneMemOperand();
3420 unsigned NewOpc;
3421 if (Instr.isBranch())
3422 NewOpc = GetCondBranchFromCond(NewCC);
3423 else if(OpcIsSET)
3424 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3425 else {
3426 unsigned DstReg = Instr.getOperand(0).getReg();
3427 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3428 HasMemoryOperand);
3429 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3430
3431 // Push the MachineInstr to OpsToUpdate.
3432 // If it is safe to remove CmpInstr, the condition code of these
3433 // instructions will be modified.
3434 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3435 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3436 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3437 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3438 IsSafe = true;
3439 break;
3440 }
3441 }
3442
3443 // If EFLAGS is not killed nor re-defined, we should check whether it is
3444 // live-out. If it is live-out, do not optimize.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3445 if ((IsCmpZero || IsSwapped) && !IsSafe) {
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3446 MachineBasicBlock *MBB = CmpInstr->getParent();
3447 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3448 SE = MBB->succ_end(); SI != SE; ++SI)
3449 if ((*SI)->isLiveIn(X86::EFLAGS))
3450 return false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3451 }
3452
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3453 // The instruction to be updated is either Sub or MI.
3454 Sub = IsCmpZero ? MI : Sub;
David Majnemer5ba473a2013-05-18 01:02:03 +0000[diff] [blame]3455 // Move Movr0Inst to the appropriate place before Sub.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3456 if (Movr0Inst) {
David Majnemer5ba473a2013-05-18 01:02:03 +0000[diff] [blame]3457 // Look backwards until we find a def that doesn't use the current EFLAGS.
3458 Def = Sub;
3459 MachineBasicBlock::reverse_iterator
3460 InsertI = MachineBasicBlock::reverse_iterator(++Def),
3461 InsertE = Sub->getParent()->rend();
3462 for (; InsertI != InsertE; ++InsertI) {
3463 MachineInstr *Instr = &*InsertI;
3464 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3465 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3466 Sub->getParent()->remove(Movr0Inst);
3467 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3468 Movr0Inst);
3469 break;
3470 }
3471 }
3472 if (InsertI == InsertE)
3473 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3474 }
3475
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3476 // Make sure Sub instruction defines EFLAGS and mark the def live.
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3477 unsigned i = 0, e = Sub->getNumOperands();
3478 for (; i != e; ++i) {
3479 MachineOperand &MO = Sub->getOperand(i);
3480 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3481 MO.setIsDead(false);
3482 break;
3483 }
3484 }
3485 assert(i != e && "Unable to locate a def EFLAGS operand");
3486
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3487 CmpInstr->eraseFromParent();
3488
3489 // Modify the condition code of instructions in OpsToUpdate.
3490 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3491 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3492 return true;
3493}
3494
Manman Ren5759d012012-08-02 00:56:42 +0000[diff] [blame]3495/// optimizeLoadInstr - Try to remove the load by folding it to a register
3496/// operand at the use. We fold the load instructions if load defines a virtual
3497/// register, the virtual register is used once in the same BB, and the
3498/// instructions in-between do not load or store, and have no side effects.
3499MachineInstr* X86InstrInfo::
3500optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3501 unsigned &FoldAsLoadDefReg,
3502 MachineInstr *&DefMI) const {
3503 if (FoldAsLoadDefReg == 0)
3504 return 0;
3505 // To be conservative, if there exists another load, clear the load candidate.
3506 if (MI->mayLoad()) {
3507 FoldAsLoadDefReg = 0;
3508 return 0;
3509 }
3510
3511 // Check whether we can move DefMI here.
3512 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3513 assert(DefMI);
3514 bool SawStore = false;
3515 if (!DefMI->isSafeToMove(this, 0, SawStore))
3516 return 0;
3517
3518 // We try to commute MI if possible.
3519 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3520 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3521 // Collect information about virtual register operands of MI.
3522 unsigned SrcOperandId = 0;
3523 bool FoundSrcOperand = false;
3524 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3525 MachineOperand &MO = MI->getOperand(i);
3526 if (!MO.isReg())
3527 continue;
3528 unsigned Reg = MO.getReg();
3529 if (Reg != FoldAsLoadDefReg)
3530 continue;
3531 // Do not fold if we have a subreg use or a def or multiple uses.
3532 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3533 return 0;
3534
3535 SrcOperandId = i;
3536 FoundSrcOperand = true;
3537 }
3538 if (!FoundSrcOperand) return 0;
3539
3540 // Check whether we can fold the def into SrcOperandId.
3541 SmallVector<unsigned, 8> Ops;
3542 Ops.push_back(SrcOperandId);
3543 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3544 if (FoldMI) {
3545 FoldAsLoadDefReg = 0;
3546 return FoldMI;
3547 }
3548
3549 if (Idx == 1) {
3550 // MI was changed but it didn't help, commute it back!
3551 commuteInstruction(MI, false);
3552 return 0;
3553 }
3554
3555 // Check whether we can commute MI and enable folding.
3556 if (MI->isCommutable()) {
3557 MachineInstr *NewMI = commuteInstruction(MI, false);
3558 // Unable to commute.
3559 if (!NewMI) return 0;
3560 if (NewMI != MI) {
3561 // New instruction. It doesn't need to be kept.
3562 NewMI->eraseFromParent();
3563 return 0;
3564 }
3565 }
3566 }
3567 return 0;
3568}
3569
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3570/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3571/// instruction with two undef reads of the register being defined. This is
3572/// used for mapping:
3573/// %xmm4 = V_SET0
3574/// to:
3575/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3576///
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3577static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3578 const MCInstrDesc &Desc) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3579 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3580 unsigned Reg = MIB->getOperand(0).getReg();
3581 MIB->setDesc(Desc);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3582
3583 // MachineInstr::addOperand() will insert explicit operands before any
3584 // implicit operands.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3585 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3586 // But we don't trust that.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3587 assert(MIB->getOperand(1).getReg() == Reg &&
3588 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3589 return true;
3590}
3591
3592bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3593 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3594 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3595 switch (MI->getOpcode()) {
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3596 case X86::SETB_C8r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3597 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3598 case X86::SETB_C16r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3599 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3600 case X86::SETB_C32r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3601 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3602 case X86::SETB_C64r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3603 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3604 case X86::V_SET0:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3605 case X86::FsFLD0SS:
3606 case X86::FsFLD0SD:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3607 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]3608 case X86::AVX_SET0:
3609 assert(HasAVX && "AVX not supported");
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3610 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3611 case X86::V_SETALLONES:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3612 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3613 case X86::AVX2_SETALLONES:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3614 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
Jakob Stoklund Olesen729abd32011-10-08 18:28:28 +0000[diff] [blame]3615 case X86::TEST8ri_NOREX:
3616 MI->setDesc(get(X86::TEST8ri));
3617 return true;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3618 }
3619 return false;
3620}
3621
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3622MachineInstr*
3623X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
Evan Cheng250e9172010-04-29 01:13:30 +0000[diff] [blame]3624 int FrameIx, uint64_t Offset,
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3625 const MDNode *MDPtr,
3626 DebugLoc DL) const {
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3627 X86AddressMode AM;
3628 AM.BaseType = X86AddressMode::FrameIndexBase;
3629 AM.Base.FrameIndex = FrameIx;
3630 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
3631 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
3632 return &*MIB;
3633}
3634
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3635static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3636 const SmallVectorImpl<MachineOperand> &MOs,
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3637 MachineInstr *MI,
3638 const TargetInstrInfo &TII) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3639 // Create the base instruction with the memory operand as the first part.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3640 // Omit the implicit operands, something BuildMI can't do.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3641 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3642 MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3643 MachineInstrBuilder MIB(MF, NewMI);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3644 unsigned NumAddrOps = MOs.size();
3645 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3646 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3647 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3648 addOffset(MIB, 0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3649
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3650 // Loop over the rest of the ri operands, converting them over.
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3651 unsigned NumOps = MI->getDesc().getNumOperands()-2;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3652 for (unsigned i = 0; i != NumOps; ++i) {
3653 MachineOperand &MO = MI->getOperand(i+2);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3654 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3655 }
3656 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3657 MachineOperand &MO = MI->getOperand(i);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3658 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3659 }
3660 return MIB;
3661}
3662
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3663static MachineInstr *FuseInst(MachineFunction &MF,
3664 unsigned Opcode, unsigned OpNo,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3665 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3666 MachineInstr *MI, const TargetInstrInfo &TII) {
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3667 // Omit the implicit operands, something BuildMI can't do.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3668 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3669 MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3670 MachineInstrBuilder MIB(MF, NewMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3671
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3672 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3673 MachineOperand &MO = MI->getOperand(i);
3674 if (i == OpNo) {
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3675 assert(MO.isReg() && "Expected to fold into reg operand!");
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3676 unsigned NumAddrOps = MOs.size();
3677 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3678 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3679 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3680 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3681 } else {
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3682 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3683 }
3684 }
3685 return MIB;
3686}
3687
3688static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3689 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3690 MachineInstr *MI) {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3691 MachineFunction &MF = *MI->getParent()->getParent();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3692 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3693
3694 unsigned NumAddrOps = MOs.size();
3695 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3696 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3697 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3698 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3699 return MIB.addImm(0);
3700}
3701
3702MachineInstr*
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3703X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3704 MachineInstr *MI, unsigned i,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3705 const SmallVectorImpl<MachineOperand> &MOs,
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3706 unsigned Size, unsigned Align) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3707 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
Preston Gurdd6be4bf2013-03-27 23:16:18 +0000[diff] [blame]3708 bool isCallRegIndirect = TM.getSubtarget<X86Subtarget>().callRegIndirect();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3709 bool isTwoAddrFold = false;
Preston Gurdd6be4bf2013-03-27 23:16:18 +0000[diff] [blame]3710
3711 // Atom favors register form of call. So, we do not fold loads into calls
3712 // when X86Subtarget is Atom.
3713 if (isCallRegIndirect &&
3714 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
3715 return NULL;
3716 }
3717
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3718 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3719 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]3720 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3721
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]3722 // FIXME: AsmPrinter doesn't know how to handle
3723 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3724 if (MI->getOpcode() == X86::ADD32ri &&
3725 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3726 return NULL;
3727
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3728 MachineInstr *NewMI = NULL;
3729 // Folding a memory location into the two-address part of a two-address
3730 // instruction is different than folding it other places. It requires
3731 // replacing the *two* registers with the memory location.
3732 if (isTwoAddr && NumOps >= 2 && i < 2 &&
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3733 MI->getOperand(0).isReg() &&
3734 MI->getOperand(1).isReg() &&
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3735 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3736 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3737 isTwoAddrFold = true;
3738 } else if (i == 0) { // If operand 0
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]3739 if (MI->getOpcode() == X86::MOV32r0) {
3740 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
3741 if (NewMI)
3742 return NewMI;
Craig Topperf9115972012-08-23 04:57:36 +0000[diff] [blame]3743 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3744
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3745 OpcodeTablePtr = &RegOp2MemOpTable0;
3746 } else if (i == 1) {
3747 OpcodeTablePtr = &RegOp2MemOpTable1;
3748 } else if (i == 2) {
3749 OpcodeTablePtr = &RegOp2MemOpTable2;
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]3750 } else if (i == 3) {
3751 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3752 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3753
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3754 // If table selected...
3755 if (OpcodeTablePtr) {
3756 // Find the Opcode to fuse
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3757 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3758 OpcodeTablePtr->find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3759 if (I != OpcodeTablePtr->end()) {
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3760 unsigned Opcode = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]3761 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3762 if (Align < MinAlign)
3763 return NULL;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3764 bool NarrowToMOV32rm = false;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3765 if (Size) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]3766 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3767 if (Size < RCSize) {
3768 // Check if it's safe to fold the load. If the size of the object is
3769 // narrower than the load width, then it's not.
3770 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3771 return NULL;
3772 // If this is a 64-bit load, but the spill slot is 32, then we can do
3773 // a 32-bit load which is implicitly zero-extended. This likely is due
3774 // to liveintervalanalysis remat'ing a load from stack slot.
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3775 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3776 return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3777 Opcode = X86::MOV32rm;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3778 NarrowToMOV32rm = true;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3779 }
3780 }
3781
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3782 if (isTwoAddrFold)
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3783 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3784 else
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3785 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3786
3787 if (NarrowToMOV32rm) {
3788 // If this is the special case where we use a MOV32rm to load a 32-bit
3789 // value and zero-extend the top bits. Change the destination register
3790 // to a 32-bit one.
3791 unsigned DstReg = NewMI->getOperand(0).getReg();
3792 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3793 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3794 X86::sub_32bit));
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3795 else
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3796 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3797 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3798 return NewMI;
3799 }
3800 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3801
3802 // No fusion
Jakob Stoklund Olesen51702ec2010-07-09 20:43:09 +0000[diff] [blame]3803 if (PrintFailedFusing && !MI->isCopy())
David Greened589daf2010-01-05 01:29:29 +0000[diff] [blame]3804 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3805 return NULL;
3806}
3807
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3808/// hasPartialRegUpdate - Return true for all instructions that only update
3809/// the first 32 or 64-bits of the destination register and leave the rest
3810/// unmodified. This can be used to avoid folding loads if the instructions
3811/// only update part of the destination register, and the non-updated part is
3812/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3813/// instructions breaks the partial register dependency and it can improve
3814/// performance. e.g.:
3815///
3816/// movss (%rdi), %xmm0
3817/// cvtss2sd %xmm0, %xmm0
3818///
3819/// Instead of
3820/// cvtss2sd (%rdi), %xmm0
3821///
Bruno Cardoso Lopes7b435682011-09-15 23:04:24 +0000[diff] [blame]3822/// FIXME: This should be turned into a TSFlags.
3823///
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3824static bool hasPartialRegUpdate(unsigned Opcode) {
3825 switch (Opcode) {
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]3826 case X86::CVTSI2SSrr:
3827 case X86::CVTSI2SS64rr:
3828 case X86::CVTSI2SDrr:
3829 case X86::CVTSI2SD64rr:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3830 case X86::CVTSD2SSrr:
3831 case X86::Int_CVTSD2SSrr:
3832 case X86::CVTSS2SDrr:
3833 case X86::Int_CVTSS2SDrr:
3834 case X86::RCPSSr:
3835 case X86::RCPSSr_Int:
3836 case X86::ROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3837 case X86::ROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3838 case X86::ROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3839 case X86::ROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3840 case X86::RSQRTSSr:
3841 case X86::RSQRTSSr_Int:
3842 case X86::SQRTSSr:
3843 case X86::SQRTSSr_Int:
3844 // AVX encoded versions
3845 case X86::VCVTSD2SSrr:
3846 case X86::Int_VCVTSD2SSrr:
3847 case X86::VCVTSS2SDrr:
3848 case X86::Int_VCVTSS2SDrr:
3849 case X86::VRCPSSr:
3850 case X86::VROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3851 case X86::VROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3852 case X86::VROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3853 case X86::VROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3854 case X86::VRSQRTSSr:
3855 case X86::VSQRTSSr:
3856 return true;
3857 }
3858
3859 return false;
3860}
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3861
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]3862/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3863/// instructions we would like before a partial register update.
3864unsigned X86InstrInfo::
3865getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3866 const TargetRegisterInfo *TRI) const {
3867 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3868 return 0;
3869
3870 // If MI is marked as reading Reg, the partial register update is wanted.
3871 const MachineOperand &MO = MI->getOperand(0);
3872 unsigned Reg = MO.getReg();
3873 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3874 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3875 return 0;
3876 } else {
3877 if (MI->readsRegister(Reg, TRI))
3878 return 0;
3879 }
3880
3881 // If any of the preceding 16 instructions are reading Reg, insert a
3882 // dependency breaking instruction. The magic number is based on a few
3883 // Nehalem experiments.
3884 return 16;
3885}
3886
3887void X86InstrInfo::
3888breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3889 const TargetRegisterInfo *TRI) const {
3890 unsigned Reg = MI->getOperand(OpNum).getReg();
3891 if (X86::VR128RegClass.contains(Reg)) {
3892 // These instructions are all floating point domain, so xorps is the best
3893 // choice.
3894 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3895 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3896 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3897 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3898 } else if (X86::VR256RegClass.contains(Reg)) {
3899 // Use vxorps to clear the full ymm register.
3900 // It wants to read and write the xmm sub-register.
3901 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3902 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3903 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3904 .addReg(Reg, RegState::ImplicitDefine);
3905 } else
3906 return;
3907 MI->addRegisterKilled(Reg, TRI, true);
3908}
3909
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3910MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3911 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3912 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3913 int FrameIndex) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3914 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3915 if (NoFusing) return NULL;
3916
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3917 // Unless optimizing for size, don't fold to avoid partial
3918 // register update stalls
Bill Wendling698e84f2012-12-30 10:32:01 +0000[diff] [blame]3919 if (!MF.getFunction()->getAttributes().
3920 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3921 hasPartialRegUpdate(MI->getOpcode()))
3922 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]3923
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3924 const MachineFrameInfo *MFI = MF.getFrameInfo();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3925 unsigned Size = MFI->getObjectSize(FrameIndex);
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3926 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3927 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3928 unsigned NewOpc = 0;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3929 unsigned RCSize = 0;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3930 switch (MI->getOpcode()) {
3931 default: return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3932 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
Dan Gohman887dd1c2010-05-18 21:42:03 +0000[diff] [blame]3933 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
3934 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
3935 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3936 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3937 // Check if it's safe to fold the load. If the size of the object is
3938 // narrower than the load width, then it's not.
3939 if (Size < RCSize)
3940 return NULL;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3941 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]3942 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3943 MI->getOperand(1).ChangeToImmediate(0);
3944 } else if (Ops.size() != 1)
3945 return NULL;
3946
3947 SmallVector<MachineOperand,4> MOs;
3948 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3949 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3950}
3951
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3952MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3953 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3954 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3955 MachineInstr *LoadMI) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3956 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3957 if (NoFusing) return NULL;
3958
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3959 // Unless optimizing for size, don't fold to avoid partial
3960 // register update stalls
Bill Wendling698e84f2012-12-30 10:32:01 +0000[diff] [blame]3961 if (!MF.getFunction()->getAttributes().
3962 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3963 hasPartialRegUpdate(MI->getOpcode()))
3964 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]3965
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]3966 // Determine the alignment of the load.
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3967 unsigned Alignment = 0;
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]3968 if (LoadMI->hasOneMemOperand())
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]3969 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3970 else
3971 switch (LoadMI->getOpcode()) {
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]3972 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]3973 case X86::AVX_SET0:
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]3974 Alignment = 32;
3975 break;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3976 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3977 case X86::V_SETALLONES:
3978 Alignment = 16;
3979 break;
3980 case X86::FsFLD0SD:
3981 Alignment = 8;
3982 break;
3983 case X86::FsFLD0SS:
3984 Alignment = 4;
3985 break;
3986 default:
Eli Friedman87ef3872011-06-10 01:13:01 +0000[diff] [blame]3987 return 0;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3988 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3989 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3990 unsigned NewOpc = 0;
3991 switch (MI->getOpcode()) {
3992 default: return NULL;
3993 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]3994 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
3995 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
3996 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3997 }
3998 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]3999 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4000 MI->getOperand(1).ChangeToImmediate(0);
4001 } else if (Ops.size() != 1)
4002 return NULL;
4003
Jakob Stoklund Olesen9c473e42010-08-11 23:08:22 +0000[diff] [blame]4004 // Make sure the subregisters match.
4005 // Otherwise we risk changing the size of the load.
4006 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4007 return NULL;
4008
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4009 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4010 switch (LoadMI->getOpcode()) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4011 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4012 case X86::V_SETALLONES:
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]4013 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4014 case X86::AVX_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4015 case X86::FsFLD0SD:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4016 case X86::FsFLD0SS: {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4017 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4018 // Create a constant-pool entry and operands to load from it.
4019
Dan Gohman772952f2010-03-09 03:01:40 +0000[diff] [blame]4020 // Medium and large mode can't fold loads this way.
4021 if (TM.getCodeModel() != CodeModel::Small &&
4022 TM.getCodeModel() != CodeModel::Kernel)
4023 return NULL;
4024
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4025 // x86-32 PIC requires a PIC base register for constant pools.
4026 unsigned PICBase = 0;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4027 if (TM.getRelocationModel() == Reloc::PIC_) {
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]4028 if (TM.getSubtarget<X86Subtarget>().is64Bit())
4029 PICBase = X86::RIP;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4030 else
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4031 // FIXME: PICBase = getGlobalBaseReg(&MF);
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]4032 // This doesn't work for several reasons.
4033 // 1. GlobalBaseReg may have been spilled.
4034 // 2. It may not be live at MI.
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4035 return NULL;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4036 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4037
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4038 // Create a constant-pool entry.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4039 MachineConstantPool &MCP = *MF.getConstantPool();
Chris Lattner229907c2011-07-18 04:54:35 +0000[diff] [blame]4040 Type *Ty;
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4041 unsigned Opc = LoadMI->getOpcode();
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4042 if (Opc == X86::FsFLD0SS)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4043 Ty = Type::getFloatTy(MF.getFunction()->getContext());
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4044 else if (Opc == X86::FsFLD0SD)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4045 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4046 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
Craig Toppera4c5a472012-01-13 06:12:41 +0000[diff] [blame]4047 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4048 else
4049 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]4050
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]4051 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]4052 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4053 Constant::getNullValue(Ty);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4054 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4055
4056 // Create operands to load from the constant pool entry.
4057 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4058 MOs.push_back(MachineOperand::CreateImm(1));
4059 MOs.push_back(MachineOperand::CreateReg(0, false));
4060 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]4061 MOs.push_back(MachineOperand::CreateReg(0, false));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4062 break;
4063 }
4064 default: {
Manman Ren5b462822012-11-27 18:09:26 +0000[diff] [blame]4065 if ((LoadMI->getOpcode() == X86::MOVSSrm ||
4066 LoadMI->getOpcode() == X86::VMOVSSrm) &&
4067 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4068 > 4)
4069 // These instructions only load 32 bits, we can't fold them if the
4070 // destination register is wider than 32 bits (4 bytes).
4071 return NULL;
4072 if ((LoadMI->getOpcode() == X86::MOVSDrm ||
4073 LoadMI->getOpcode() == X86::VMOVSDrm) &&
4074 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4075 > 8)
4076 // These instructions only load 64 bits, we can't fold them if the
4077 // destination register is wider than 64 bits (8 bytes).
4078 return NULL;
4079
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4080 // Folding a normal load. Just copy the load's address operands.
4081 unsigned NumOps = LoadMI->getDesc().getNumOperands();
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4082 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4083 MOs.push_back(LoadMI->getOperand(i));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4084 break;
4085 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4086 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4087 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4088}
4089
4090
Dan Gohman33332bc2008-10-16 01:49:15 +0000[diff] [blame]4091bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4092 const SmallVectorImpl<unsigned> &Ops) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4093 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4094 if (NoFusing) return 0;
4095
4096 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4097 switch (MI->getOpcode()) {
4098 default: return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4099 case X86::TEST8rr:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4100 case X86::TEST16rr:
4101 case X86::TEST32rr:
4102 case X86::TEST64rr:
4103 return true;
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]4104 case X86::ADD32ri:
4105 // FIXME: AsmPrinter doesn't know how to handle
4106 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4107 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4108 return false;
4109 break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4110 }
4111 }
4112
4113 if (Ops.size() != 1)
4114 return false;
4115
4116 unsigned OpNum = Ops[0];
4117 unsigned Opc = MI->getOpcode();
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]4118 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4119 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4120 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4121
4122 // Folding a memory location into the two-address part of a two-address
4123 // instruction is different than folding it other places. It requires
4124 // replacing the *two* registers with the memory location.
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4125 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4126 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4127 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4128 } else if (OpNum == 0) { // If operand 0
Tim Northover64ec0ff2013-05-30 13:19:42 +0000[diff] [blame]4129 if (Opc == X86::MOV32r0)
4130 return true;
4131
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4132 OpcodeTablePtr = &RegOp2MemOpTable0;
4133 } else if (OpNum == 1) {
4134 OpcodeTablePtr = &RegOp2MemOpTable1;
4135 } else if (OpNum == 2) {
4136 OpcodeTablePtr = &RegOp2MemOpTable2;
Craig Topper7573c8f2012-08-31 22:12:16 +0000[diff] [blame]4137 } else if (OpNum == 3) {
4138 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4139 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4140
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]4141 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4142 return true;
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]4143 return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4144}
4145
4146bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4147 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4148 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4149 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4150 MemOp2RegOpTable.find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4151 if (I == MemOp2RegOpTable.end())
4152 return false;
4153 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4154 unsigned Index = I->second.second & TB_INDEX_MASK;
4155 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4156 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4157 if (UnfoldLoad && !FoldedLoad)
4158 return false;
4159 UnfoldLoad &= FoldedLoad;
4160 if (UnfoldStore && !FoldedStore)
4161 return false;
4162 UnfoldStore &= FoldedStore;
4163
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4164 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4165 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4166 if (!MI->hasOneMemOperand() &&
4167 RC == &X86::VR128RegClass &&
4168 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4169 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4170 // conservatively assume the address is unaligned. That's bad for
4171 // performance.
4172 return false;
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4173 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4174 SmallVector<MachineOperand,2> BeforeOps;
4175 SmallVector<MachineOperand,2> AfterOps;
4176 SmallVector<MachineOperand,4> ImpOps;
4177 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4178 MachineOperand &Op = MI->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4179 if (i >= Index && i < Index + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4180 AddrOps.push_back(Op);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4181 else if (Op.isReg() && Op.isImplicit())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4182 ImpOps.push_back(Op);
4183 else if (i < Index)
4184 BeforeOps.push_back(Op);
4185 else if (i > Index)
4186 AfterOps.push_back(Op);
4187 }
4188
4189 // Emit the load instruction.
4190 if (UnfoldLoad) {
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4191 std::pair<MachineInstr::mmo_iterator,
4192 MachineInstr::mmo_iterator> MMOs =
4193 MF.extractLoadMemRefs(MI->memoperands_begin(),
4194 MI->memoperands_end());
4195 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4196 if (UnfoldStore) {
4197 // Address operands cannot be marked isKill.
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4198 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4199 MachineOperand &MO = NewMIs[0]->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4200 if (MO.isReg())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4201 MO.setIsKill(false);
4202 }
4203 }
4204 }
4205
4206 // Emit the data processing instruction.
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4207 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]4208 MachineInstrBuilder MIB(MF, DataMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4209
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4210 if (FoldedStore)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4211 MIB.addReg(Reg, RegState::Define);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4212 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4213 MIB.addOperand(BeforeOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4214 if (FoldedLoad)
4215 MIB.addReg(Reg);
4216 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4217 MIB.addOperand(AfterOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4218 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4219 MachineOperand &MO = ImpOps[i];
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4220 MIB.addReg(MO.getReg(),
4221 getDefRegState(MO.isDef()) |
4222 RegState::Implicit |
4223 getKillRegState(MO.isKill()) |
Evan Cheng0dc101b2009-06-30 08:49:04 +0000[diff] [blame]4224 getDeadRegState(MO.isDead()) |
4225 getUndefRegState(MO.isUndef()));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4226 }
4227 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4228 switch (DataMI->getOpcode()) {
4229 default: break;
4230 case X86::CMP64ri32:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4231 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4232 case X86::CMP32ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4233 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4234 case X86::CMP16ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4235 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4236 case X86::CMP8ri: {
4237 MachineOperand &MO0 = DataMI->getOperand(0);
4238 MachineOperand &MO1 = DataMI->getOperand(1);
4239 if (MO1.getImm() == 0) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4240 unsigned NewOpc;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4241 switch (DataMI->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4242 default: llvm_unreachable("Unreachable!");
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4243 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4244 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4245 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4246 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4247 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4248 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4249 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4250 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4251 DataMI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4252 MO1.ChangeToRegister(MO0.getReg(), false);
4253 }
4254 }
4255 }
4256 NewMIs.push_back(DataMI);
4257
4258 // Emit the store instruction.
4259 if (UnfoldStore) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4260 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4261 std::pair<MachineInstr::mmo_iterator,
4262 MachineInstr::mmo_iterator> MMOs =
4263 MF.extractStoreMemRefs(MI->memoperands_begin(),
4264 MI->memoperands_end());
4265 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4266 }
4267
4268 return true;
4269}
4270
4271bool
4272X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4273 SmallVectorImpl<SDNode*> &NewNodes) const {
Dan Gohman17059682008-07-17 19:10:17 +0000[diff] [blame]4274 if (!N->isMachineOpcode())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4275 return false;
4276
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4277 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4278 MemOp2RegOpTable.find(N->getMachineOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4279 if (I == MemOp2RegOpTable.end())
4280 return false;
4281 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4282 unsigned Index = I->second.second & TB_INDEX_MASK;
4283 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4284 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4285 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4286 MachineFunction &MF = DAG.getMachineFunction();
4287 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4288 unsigned NumDefs = MCID.NumDefs;
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4289 std::vector<SDValue> AddrOps;
4290 std::vector<SDValue> BeforeOps;
4291 std::vector<SDValue> AfterOps;
Andrew Trickef9de2a2013-05-25 02:42:55 +0000[diff] [blame]4292 SDLoc dl(N);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4293 unsigned NumOps = N->getNumOperands();
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]4294 for (unsigned i = 0; i != NumOps-1; ++i) {
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4295 SDValue Op = N->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4296 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4297 AddrOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4298 else if (i < Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4299 BeforeOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4300 else if (i > Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4301 AfterOps.push_back(Op);
4302 }
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4303 SDValue Chain = N->getOperand(NumOps-1);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4304 AddrOps.push_back(Chain);
4305
4306 // Emit the load instruction.
4307 SDNode *Load = 0;
4308 if (FoldedLoad) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4309 EVT VT = *RC->vt_begin();
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4310 std::pair<MachineInstr::mmo_iterator,
4311 MachineInstr::mmo_iterator> MMOs =
4312 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4313 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4314 if (!(*MMOs.first) &&
4315 RC == &X86::VR128RegClass &&
4316 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4317 // Do not introduce a slow unaligned load.
4318 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4319 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4320 bool isAligned = (*MMOs.first) &&
4321 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4322 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4323 VT, MVT::Other, AddrOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4324 NewNodes.push_back(Load);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4325
4326 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4327 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4328 }
4329
4330 // Emit the data processing instruction.
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4331 std::vector<EVT> VTs;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4332 const TargetRegisterClass *DstRC = 0;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4333 if (MCID.getNumDefs() > 0) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4334 DstRC = getRegClass(MCID, 0, &RI, MF);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4335 VTs.push_back(*DstRC->vt_begin());
4336 }
4337 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4338 EVT VT = N->getValueType(i);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4339 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4340 VTs.push_back(VT);
4341 }
4342 if (Load)
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4343 BeforeOps.push_back(SDValue(Load, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4344 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4345 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4346 NewNodes.push_back(NewNode);
4347
4348 // Emit the store instruction.
4349 if (FoldedStore) {
4350 AddrOps.pop_back();
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4351 AddrOps.push_back(SDValue(NewNode, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4352 AddrOps.push_back(Chain);
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4353 std::pair<MachineInstr::mmo_iterator,
4354 MachineInstr::mmo_iterator> MMOs =
4355 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4356 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4357 if (!(*MMOs.first) &&
4358 RC == &X86::VR128RegClass &&
4359 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4360 // Do not introduce a slow unaligned store.
4361 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4362 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4363 bool isAligned = (*MMOs.first) &&
4364 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4365 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4366 isAligned, TM),
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4367 dl, MVT::Other, AddrOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4368 NewNodes.push_back(Store);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4369
4370 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4371 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4372 }
4373
4374 return true;
4375}
4376
4377unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4378 bool UnfoldLoad, bool UnfoldStore,
4379 unsigned *LoadRegIndex) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4380 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4381 MemOp2RegOpTable.find(Opc);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4382 if (I == MemOp2RegOpTable.end())
4383 return 0;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4384 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4385 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4386 if (UnfoldLoad && !FoldedLoad)
4387 return 0;
4388 if (UnfoldStore && !FoldedStore)
4389 return 0;
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4390 if (LoadRegIndex)
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4391 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4392 return I->second.first;
4393}
4394
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4395bool
4396X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4397 int64_t &Offset1, int64_t &Offset2) const {
4398 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4399 return false;
4400 unsigned Opc1 = Load1->getMachineOpcode();
4401 unsigned Opc2 = Load2->getMachineOpcode();
4402 switch (Opc1) {
4403 default: return false;
4404 case X86::MOV8rm:
4405 case X86::MOV16rm:
4406 case X86::MOV32rm:
4407 case X86::MOV64rm:
4408 case X86::LD_Fp32m:
4409 case X86::LD_Fp64m:
4410 case X86::LD_Fp80m:
4411 case X86::MOVSSrm:
4412 case X86::MOVSDrm:
4413 case X86::MMX_MOVD64rm:
4414 case X86::MMX_MOVQ64rm:
4415 case X86::FsMOVAPSrm:
4416 case X86::FsMOVAPDrm:
4417 case X86::MOVAPSrm:
4418 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4419 case X86::MOVAPDrm:
4420 case X86::MOVDQArm:
4421 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4422 // AVX load instructions
4423 case X86::VMOVSSrm:
4424 case X86::VMOVSDrm:
4425 case X86::FsVMOVAPSrm:
4426 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4427 case X86::VMOVAPSrm:
4428 case X86::VMOVUPSrm:
4429 case X86::VMOVAPDrm:
4430 case X86::VMOVDQArm:
4431 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4432 case X86::VMOVAPSYrm:
4433 case X86::VMOVUPSYrm:
4434 case X86::VMOVAPDYrm:
4435 case X86::VMOVDQAYrm:
4436 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4437 break;
4438 }
4439 switch (Opc2) {
4440 default: return false;
4441 case X86::MOV8rm:
4442 case X86::MOV16rm:
4443 case X86::MOV32rm:
4444 case X86::MOV64rm:
4445 case X86::LD_Fp32m:
4446 case X86::LD_Fp64m:
4447 case X86::LD_Fp80m:
4448 case X86::MOVSSrm:
4449 case X86::MOVSDrm:
4450 case X86::MMX_MOVD64rm:
4451 case X86::MMX_MOVQ64rm:
4452 case X86::FsMOVAPSrm:
4453 case X86::FsMOVAPDrm:
4454 case X86::MOVAPSrm:
4455 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4456 case X86::MOVAPDrm:
4457 case X86::MOVDQArm:
4458 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4459 // AVX load instructions
4460 case X86::VMOVSSrm:
4461 case X86::VMOVSDrm:
4462 case X86::FsVMOVAPSrm:
4463 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4464 case X86::VMOVAPSrm:
4465 case X86::VMOVUPSrm:
4466 case X86::VMOVAPDrm:
4467 case X86::VMOVDQArm:
4468 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4469 case X86::VMOVAPSYrm:
4470 case X86::VMOVUPSYrm:
4471 case X86::VMOVAPDYrm:
4472 case X86::VMOVDQAYrm:
4473 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4474 break;
4475 }
4476
4477 // Check if chain operands and base addresses match.
4478 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4479 Load1->getOperand(5) != Load2->getOperand(5))
4480 return false;
4481 // Segment operands should match as well.
4482 if (Load1->getOperand(4) != Load2->getOperand(4))
4483 return false;
4484 // Scale should be 1, Index should be Reg0.
4485 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4486 Load1->getOperand(2) == Load2->getOperand(2)) {
4487 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4488 return false;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4489
4490 // Now let's examine the displacements.
4491 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4492 isa<ConstantSDNode>(Load2->getOperand(3))) {
4493 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4494 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4495 return true;
4496 }
4497 }
4498 return false;
4499}
4500
4501bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4502 int64_t Offset1, int64_t Offset2,
4503 unsigned NumLoads) const {
4504 assert(Offset2 > Offset1);
4505 if ((Offset2 - Offset1) / 8 > 64)
4506 return false;
4507
4508 unsigned Opc1 = Load1->getMachineOpcode();
4509 unsigned Opc2 = Load2->getMachineOpcode();
4510 if (Opc1 != Opc2)
4511 return false; // FIXME: overly conservative?
4512
4513 switch (Opc1) {
4514 default: break;
4515 case X86::LD_Fp32m:
4516 case X86::LD_Fp64m:
4517 case X86::LD_Fp80m:
4518 case X86::MMX_MOVD64rm:
4519 case X86::MMX_MOVQ64rm:
4520 return false;
4521 }
4522
4523 EVT VT = Load1->getValueType(0);
4524 switch (VT.getSimpleVT().SimpleTy) {
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4525 default:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4526 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4527 // have 16 of them to play with.
4528 if (TM.getSubtargetImpl()->is64Bit()) {
4529 if (NumLoads >= 3)
4530 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4531 } else if (NumLoads) {
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4532 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4533 }
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4534 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4535 case MVT::i8:
4536 case MVT::i16:
4537 case MVT::i32:
4538 case MVT::i64:
Evan Cheng16cf9342010-01-22 23:49:11 +0000[diff] [blame]4539 case MVT::f32:
4540 case MVT::f64:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4541 if (NumLoads)
4542 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4543 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4544 }
4545
4546 return true;
4547}
4548
4549
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4550bool X86InstrInfo::
Owen Anderson4f6bf042008-08-14 22:49:33 +0000[diff] [blame]4551ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4552 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4553 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]4554 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4555 return true;
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4556 Cond[0].setImm(GetOppositeBranchCondition(CC));
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4557 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4558}
4559
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4560bool X86InstrInfo::
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4561isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4562 // FIXME: Return false for x87 stack register classes for now. We can't
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4563 // allow any loads of these registers before FpGet_ST0_80.
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4564 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4565 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4566}
4567
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4568/// getGlobalBaseReg - Return a virtual register initialized with the
4569/// the global base register value. Output instructions required to
4570/// initialize the register in the function entry block, if necessary.
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4571///
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4572/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4573///
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4574unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4575 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4576 "X86-64 PIC uses RIP relative addressing");
4577
4578 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4579 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4580 if (GlobalBaseReg != 0)
4581 return GlobalBaseReg;
4582
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4583 // Create the register. The code to initialize it is inserted
4584 // later, by the CGBR pass (below).
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4585 MachineRegisterInfo &RegInfo = MF->getRegInfo();
Jakob Stoklund Olesen38dcd592012-05-20 18:43:00 +0000[diff] [blame]4586 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4587 X86FI->setGlobalBaseReg(GlobalBaseReg);
4588 return GlobalBaseReg;
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4589}
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4590
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4591// These are the replaceable SSE instructions. Some of these have Int variants
4592// that we don't include here. We don't want to replace instructions selected
4593// by intrinsics.
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4594static const uint16_t ReplaceableInstrs[][3] = {
Bruno Cardoso Lopes1401e042010-08-12 02:08:52 +0000[diff] [blame]4595 //PackedSingle PackedDouble PackedInt
Jakob Stoklund Olesendbff4e82010-03-30 22:46:53 +0000[diff] [blame]4596 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4597 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4598 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4599 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4600 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4601 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4602 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4603 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4604 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4605 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4606 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4607 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4608 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4609 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4610 // AVX 128-bit support
4611 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4612 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4613 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4614 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4615 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4616 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4617 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4618 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4619 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4620 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4621 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4622 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4623 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4624 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4625 // AVX 256-bit support
4626 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4627 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4628 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4629 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4630 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4631 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4632};
4633
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4634static const uint16_t ReplaceableInstrsAVX2[][3] = {
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4635 //PackedSingle PackedDouble PackedInt
Craig Topperf87a2be2011-11-09 09:37:21 +0000[diff] [blame]4636 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4637 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4638 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4639 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4640 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4641 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4642 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
Craig Topper12b72de2011-11-29 05:37:58 +0000[diff] [blame]4643 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4644 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4645 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4646 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4647 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4648 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4649 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4650};
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4651
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4652// FIXME: Some shuffle and unpack instructions have equivalents in different
4653// domains, but they require a bit more work than just switching opcodes.
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4654
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4655static const uint16_t *lookup(unsigned opcode, unsigned domain) {
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4656 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4657 if (ReplaceableInstrs[i][domain-1] == opcode)
4658 return ReplaceableInstrs[i];
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4659 return 0;
4660}
4661
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4662static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4663 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4664 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4665 return ReplaceableInstrsAVX2[i];
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4666 return 0;
4667}
4668
4669std::pair<uint16_t, uint16_t>
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]4670X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4671 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4672 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4673 uint16_t validDomains = 0;
4674 if (domain && lookup(MI->getOpcode(), domain))
4675 validDomains = 0xe;
4676 else if (domain && lookupAVX2(MI->getOpcode(), domain))
4677 validDomains = hasAVX2 ? 0xe : 0x6;
4678 return std::make_pair(domain, validDomains);
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4679}
4680
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]4681void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4682 assert(Domain>0 && Domain<4 && "Invalid execution domain");
4683 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4684 assert(dom && "Not an SSE instruction");
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4685 const uint16_t *table = lookup(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]4686 if (!table) { // try the other table
4687 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
4688 "256-bit vector operations only available in AVX2");
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4689 table = lookupAVX2(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]4690 }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4691 assert(table && "Cannot change domain");
4692 MI->setDesc(get(table[Domain-1]));
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4693}
Chris Lattner6a5e7062010-04-26 23:37:21 +0000[diff] [blame]4694
4695/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
4696void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
4697 NopInst.setOpcode(X86::NOOP);
4698}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4699
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]4700bool X86InstrInfo::isHighLatencyDef(int opc) const {
4701 switch (opc) {
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4702 default: return false;
4703 case X86::DIVSDrm:
4704 case X86::DIVSDrm_Int:
4705 case X86::DIVSDrr:
4706 case X86::DIVSDrr_Int:
4707 case X86::DIVSSrm:
4708 case X86::DIVSSrm_Int:
4709 case X86::DIVSSrr:
4710 case X86::DIVSSrr_Int:
4711 case X86::SQRTPDm:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4712 case X86::SQRTPDr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4713 case X86::SQRTPSm:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4714 case X86::SQRTPSr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4715 case X86::SQRTSDm:
4716 case X86::SQRTSDm_Int:
4717 case X86::SQRTSDr:
4718 case X86::SQRTSDr_Int:
4719 case X86::SQRTSSm:
4720 case X86::SQRTSSm_Int:
4721 case X86::SQRTSSr:
4722 case X86::SQRTSSr_Int:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4723 // AVX instructions with high latency
4724 case X86::VDIVSDrm:
4725 case X86::VDIVSDrm_Int:
4726 case X86::VDIVSDrr:
4727 case X86::VDIVSDrr_Int:
4728 case X86::VDIVSSrm:
4729 case X86::VDIVSSrm_Int:
4730 case X86::VDIVSSrr:
4731 case X86::VDIVSSrr_Int:
4732 case X86::VSQRTPDm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4733 case X86::VSQRTPDr:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4734 case X86::VSQRTPSm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4735 case X86::VSQRTPSr:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4736 case X86::VSQRTSDm:
4737 case X86::VSQRTSDm_Int:
4738 case X86::VSQRTSDr:
4739 case X86::VSQRTSSm:
4740 case X86::VSQRTSSm_Int:
4741 case X86::VSQRTSSr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4742 return true;
4743 }
4744}
4745
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]4746bool X86InstrInfo::
4747hasHighOperandLatency(const InstrItineraryData *ItinData,
4748 const MachineRegisterInfo *MRI,
4749 const MachineInstr *DefMI, unsigned DefIdx,
4750 const MachineInstr *UseMI, unsigned UseIdx) const {
4751 return isHighLatencyDef(DefMI->getOpcode());
4752}
4753
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4754namespace {
4755 /// CGBR - Create Global Base Reg pass. This initializes the PIC
4756 /// global base register for x86-32.
4757 struct CGBR : public MachineFunctionPass {
4758 static char ID;
Owen Andersona7aed182010-08-06 18:33:48 +0000[diff] [blame]4759 CGBR() : MachineFunctionPass(ID) {}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4760
4761 virtual bool runOnMachineFunction(MachineFunction &MF) {
4762 const X86TargetMachine *TM =
4763 static_cast<const X86TargetMachine *>(&MF.getTarget());
4764
4765 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
4766 "X86-64 PIC uses RIP relative addressing");
4767
4768 // Only emit a global base reg in PIC mode.
4769 if (TM->getRelocationModel() != Reloc::PIC_)
4770 return false;
4771
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]4772 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
4773 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4774
4775 // If we didn't need a GlobalBaseReg, don't insert code.
4776 if (GlobalBaseReg == 0)
4777 return false;
4778
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4779 // Insert the set of GlobalBaseReg into the first MBB of the function
4780 MachineBasicBlock &FirstMBB = MF.front();
4781 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
4782 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
4783 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4784 const X86InstrInfo *TII = TM->getInstrInfo();
4785
4786 unsigned PC;
4787 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]4788 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4789 else
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]4790 PC = GlobalBaseReg;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4791
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4792 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4793 // only used in JIT code emission as displacement to pc.
4794 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4795
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4796 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4797 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4798 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4799 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4800 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4801 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4802 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4803 }
4804
4805 return true;
4806 }
4807
4808 virtual const char *getPassName() const {
4809 return "X86 PIC Global Base Reg Initialization";
4810 }
4811
4812 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4813 AU.setPreservesCFG();
4814 MachineFunctionPass::getAnalysisUsage(AU);
4815 }
4816 };
4817}
4818
4819char CGBR::ID = 0;
4820FunctionPass*
4821llvm::createGlobalBaseRegPass() { return new CGBR(); }
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]4822
4823namespace {
4824 struct LDTLSCleanup : public MachineFunctionPass {
4825 static char ID;
4826 LDTLSCleanup() : MachineFunctionPass(ID) {}
4827
4828 virtual bool runOnMachineFunction(MachineFunction &MF) {
4829 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4830 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4831 // No point folding accesses if there isn't at least two.
4832 return false;
4833 }
4834
4835 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4836 return VisitNode(DT->getRootNode(), 0);
4837 }
4838
4839 // Visit the dominator subtree rooted at Node in pre-order.
4840 // If TLSBaseAddrReg is non-null, then use that to replace any
4841 // TLS_base_addr instructions. Otherwise, create the register
4842 // when the first such instruction is seen, and then use it
4843 // as we encounter more instructions.
4844 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4845 MachineBasicBlock *BB = Node->getBlock();
4846 bool Changed = false;
4847
4848 // Traverse the current block.
4849 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4850 ++I) {
4851 switch (I->getOpcode()) {
4852 case X86::TLS_base_addr32:
4853 case X86::TLS_base_addr64:
4854 if (TLSBaseAddrReg)
4855 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4856 else
4857 I = SetRegister(I, &TLSBaseAddrReg);
4858 Changed = true;
4859 break;
4860 default:
4861 break;
4862 }
4863 }
4864
4865 // Visit the children of this block in the dominator tree.
4866 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4867 I != E; ++I) {
4868 Changed |= VisitNode(*I, TLSBaseAddrReg);
4869 }
4870
4871 return Changed;
4872 }
4873
4874 // Replace the TLS_base_addr instruction I with a copy from
4875 // TLSBaseAddrReg, returning the new instruction.
4876 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4877 unsigned TLSBaseAddrReg) {
4878 MachineFunction *MF = I->getParent()->getParent();
4879 const X86TargetMachine *TM =
4880 static_cast<const X86TargetMachine *>(&MF->getTarget());
4881 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4882 const X86InstrInfo *TII = TM->getInstrInfo();
4883
4884 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4885 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4886 TII->get(TargetOpcode::COPY),
4887 is64Bit ? X86::RAX : X86::EAX)
4888 .addReg(TLSBaseAddrReg);
4889
4890 // Erase the TLS_base_addr instruction.
4891 I->eraseFromParent();
4892
4893 return Copy;
4894 }
4895
4896 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4897 // inserting a copy instruction after I. Returns the new instruction.
4898 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4899 MachineFunction *MF = I->getParent()->getParent();
4900 const X86TargetMachine *TM =
4901 static_cast<const X86TargetMachine *>(&MF->getTarget());
4902 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4903 const X86InstrInfo *TII = TM->getInstrInfo();
4904
4905 // Create a virtual register for the TLS base address.
4906 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4907 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4908 ? &X86::GR64RegClass
4909 : &X86::GR32RegClass);
4910
4911 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4912 MachineInstr *Next = I->getNextNode();
4913 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4914 TII->get(TargetOpcode::COPY),
4915 *TLSBaseAddrReg)
4916 .addReg(is64Bit ? X86::RAX : X86::EAX);
4917
4918 return Copy;
4919 }
4920
4921 virtual const char *getPassName() const {
4922 return "Local Dynamic TLS Access Clean-up";
4923 }
4924
4925 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4926 AU.setPreservesCFG();
4927 AU.addRequired<MachineDominatorTree>();
4928 MachineFunctionPass::getAnalysisUsage(AU);
4929 }
4930 };
4931}
4932
4933char LDTLSCleanup::ID = 0;
4934FunctionPass*
4935llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }