Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / 4bc3e5a1bfc36eac11750c70eaaba33e79acd85c / . / llvm / lib / Target / X86 / X86InstrInfo.cpp
blob: cc35b4a02d7065c24b921c81cdbfb251bec61e4c [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"
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]20#include "llvm/DerivedTypes.h"
Owen Anderson53a52212009-07-13 04:09:18 +0000[diff] [blame]21#include "llvm/LLVMContext.h"
Owen Andersone2f23a32007-09-07 04:06:50 +0000[diff] [blame]22#include "llvm/ADT/STLExtras.h"
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]23#include "llvm/CodeGen/MachineConstantPool.h"
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]24#include "llvm/CodeGen/MachineDominators.h"
Owen Anderson6bb0c522008-01-04 23:57:37 +0000[diff] [blame]25#include "llvm/CodeGen/MachineFrameInfo.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]26#include "llvm/CodeGen/MachineInstrBuilder.h"
Chris Lattnera10fff52007-12-31 04:13:23 +0000[diff] [blame]27#include "llvm/CodeGen/MachineRegisterInfo.h"
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]28#include "llvm/CodeGen/LiveVariables.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 },
300 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
301 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
302 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Bruno Cardoso 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
358 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
359 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
360 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Craig 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 },
454 { X86::MOVZX64rr16, X86::MOVZX64rm16, 0 },
455 { X86::MOVZX64rr32, X86::MOVZX64rm32, 0 },
456 { X86::MOVZX64rr8, X86::MOVZX64rm8, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]457 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
458 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
459 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]460 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
461 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
462 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
463 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
464 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
465 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
466 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
467 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
468 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
469 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
470 { X86::SQRTPDr_Int, X86::SQRTPDm_Int, TB_ALIGN_16 },
471 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
472 { X86::SQRTPSr_Int, X86::SQRTPSm_Int, TB_ALIGN_16 },
473 { X86::SQRTSDr, X86::SQRTSDm, 0 },
474 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
475 { X86::SQRTSSr, X86::SQRTSSm, 0 },
476 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
477 { X86::TEST16rr, X86::TEST16rm, 0 },
478 { X86::TEST32rr, X86::TEST32rm, 0 },
479 { X86::TEST64rr, X86::TEST64rm, 0 },
480 { X86::TEST8rr, X86::TEST8rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]481 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]482 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
483 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]484 // AVX 128-bit versions of foldable instructions
485 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
486 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]487 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
488 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]489 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
490 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
Pete Cooper8bbce762012-06-14 22:12:58 +0000[diff] [blame]491 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]492 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
493 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
494 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
495 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
496 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
497 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
498 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
499 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
500 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]501 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
502 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
503 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
504 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
505 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
506 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
507 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
508 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
509 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
510 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
511 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
512 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
513 { X86::VMOVUPDrr, X86::VMOVUPDrm, TB_ALIGN_16 },
514 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
515 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
516 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
517 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]518 { X86::VPABSBrr128, X86::VPABSBrm128, TB_ALIGN_16 },
519 { X86::VPABSDrr128, X86::VPABSDrm128, TB_ALIGN_16 },
520 { X86::VPABSWrr128, X86::VPABSWrm128, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]521 { X86::VPERMILPDri, X86::VPERMILPDmi, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]522 { X86::VPERMILPSri, X86::VPERMILPSmi, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]523 { X86::VPSHUFDri, X86::VPSHUFDmi, TB_ALIGN_16 },
524 { X86::VPSHUFHWri, X86::VPSHUFHWmi, TB_ALIGN_16 },
525 { X86::VPSHUFLWri, X86::VPSHUFLWmi, TB_ALIGN_16 },
526 { X86::VRCPPSr, X86::VRCPPSm, TB_ALIGN_16 },
527 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, TB_ALIGN_16 },
528 { X86::VRSQRTPSr, X86::VRSQRTPSm, TB_ALIGN_16 },
529 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, TB_ALIGN_16 },
530 { X86::VSQRTPDr, X86::VSQRTPDm, TB_ALIGN_16 },
531 { X86::VSQRTPDr_Int, X86::VSQRTPDm_Int, TB_ALIGN_16 },
532 { X86::VSQRTPSr, X86::VSQRTPSm, TB_ALIGN_16 },
533 { X86::VSQRTPSr_Int, X86::VSQRTPSm_Int, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]534 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]535 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]536 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
537
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]538 // AVX 256-bit foldable instructions
539 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
540 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]541 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]542 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]543 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]544 { X86::VPERMILPDYri, X86::VPERMILPDYmi, TB_ALIGN_32 },
545 { X86::VPERMILPSYri, X86::VPERMILPSYmi, TB_ALIGN_32 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]546
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]547 // AVX2 foldable instructions
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]548 { X86::VPABSBrr256, X86::VPABSBrm256, TB_ALIGN_32 },
549 { X86::VPABSDrr256, X86::VPABSDrm256, TB_ALIGN_32 },
550 { X86::VPABSWrr256, X86::VPABSWrm256, TB_ALIGN_32 },
551 { X86::VPSHUFDYri, X86::VPSHUFDYmi, TB_ALIGN_32 },
552 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, TB_ALIGN_32 },
553 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, TB_ALIGN_32 },
554 { X86::VRCPPSYr, X86::VRCPPSYm, TB_ALIGN_32 },
555 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, TB_ALIGN_32 },
556 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, TB_ALIGN_32 },
557 { X86::VRSQRTPSYr_Int, X86::VRSQRTPSYm_Int, TB_ALIGN_32 },
558 { X86::VSQRTPDYr, X86::VSQRTPDYm, TB_ALIGN_32 },
559 { X86::VSQRTPDYr_Int, X86::VSQRTPDYm_Int, TB_ALIGN_32 },
560 { X86::VSQRTPSYr, X86::VSQRTPSYm, TB_ALIGN_32 },
561 { X86::VSQRTPSYr_Int, X86::VSQRTPSYm_Int, TB_ALIGN_32 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]562 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
563 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]564 };
565
566 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]567 unsigned RegOp = OpTbl1[i].RegOp;
568 unsigned MemOp = OpTbl1[i].MemOp;
569 unsigned Flags = OpTbl1[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]570 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
571 RegOp, MemOp,
572 // Index 1, folded load
573 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]574 }
575
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]576 static const X86OpTblEntry OpTbl2[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]577 { X86::ADC32rr, X86::ADC32rm, 0 },
578 { X86::ADC64rr, X86::ADC64rm, 0 },
579 { X86::ADD16rr, X86::ADD16rm, 0 },
580 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
581 { X86::ADD32rr, X86::ADD32rm, 0 },
582 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
583 { X86::ADD64rr, X86::ADD64rm, 0 },
584 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
585 { X86::ADD8rr, X86::ADD8rm, 0 },
586 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
587 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
588 { X86::ADDSDrr, X86::ADDSDrm, 0 },
589 { X86::ADDSSrr, X86::ADDSSrm, 0 },
590 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
591 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
592 { X86::AND16rr, X86::AND16rm, 0 },
593 { X86::AND32rr, X86::AND32rm, 0 },
594 { X86::AND64rr, X86::AND64rm, 0 },
595 { X86::AND8rr, X86::AND8rm, 0 },
596 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
597 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
598 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
599 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]600 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
601 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
602 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
603 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]604 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
605 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
606 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
607 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
608 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
609 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
610 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
611 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
612 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
613 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
614 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
615 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
616 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
617 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
618 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
619 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
620 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
621 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
622 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
623 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
624 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
625 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
626 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
627 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
628 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
629 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
630 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
631 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
632 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
633 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
634 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
635 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
636 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
637 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
638 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
639 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
640 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
641 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
642 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
643 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
644 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
645 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
646 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
647 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
648 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
649 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
650 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
651 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
652 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
653 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
654 { X86::CMPSDrr, X86::CMPSDrm, 0 },
655 { X86::CMPSSrr, X86::CMPSSrm, 0 },
656 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
657 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
658 { X86::DIVSDrr, X86::DIVSDrm, 0 },
659 { X86::DIVSSrr, X86::DIVSSrm, 0 },
660 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
661 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
662 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
663 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
664 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
665 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
666 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
667 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
668 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
669 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
670 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
671 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
672 { X86::IMUL16rr, X86::IMUL16rm, 0 },
673 { X86::IMUL32rr, X86::IMUL32rm, 0 },
674 { X86::IMUL64rr, X86::IMUL64rm, 0 },
675 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
676 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
Manman Ren959acb12012-08-13 18:29:41 +0000[diff] [blame]677 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
678 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
679 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
680 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
681 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
682 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]683 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
684 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, TB_ALIGN_16 },
685 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
686 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, TB_ALIGN_16 },
687 { X86::MAXSDrr, X86::MAXSDrm, 0 },
688 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
689 { X86::MAXSSrr, X86::MAXSSrm, 0 },
690 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
691 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
692 { X86::MINPDrr_Int, X86::MINPDrm_Int, TB_ALIGN_16 },
693 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
694 { X86::MINPSrr_Int, X86::MINPSrm_Int, TB_ALIGN_16 },
695 { X86::MINSDrr, X86::MINSDrm, 0 },
696 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
697 { X86::MINSSrr, X86::MINSSrm, 0 },
698 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]699 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]700 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
701 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
702 { X86::MULSDrr, X86::MULSDrm, 0 },
703 { X86::MULSSrr, X86::MULSSrm, 0 },
704 { X86::OR16rr, X86::OR16rm, 0 },
705 { X86::OR32rr, X86::OR32rm, 0 },
706 { X86::OR64rr, X86::OR64rm, 0 },
707 { X86::OR8rr, X86::OR8rm, 0 },
708 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
709 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
710 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
711 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]712 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]713 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
714 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
715 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
716 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
717 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
718 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]719 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
720 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]721 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]722 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]723 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
724 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
725 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
726 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]727 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]728 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
729 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]730 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]731 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
732 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
733 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]734 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]735 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]736 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
737 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]738 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]739 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]740 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]741 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]742 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]743 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]744 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
745 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
746 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
747 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
748 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
749 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]750 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]751 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
752 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
753 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
754 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
755 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
756 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
757 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]758 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
759 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
760 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
761 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]762 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
763 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
764 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
765 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
766 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
767 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
768 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
769 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
770 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
771 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
772 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
773 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
774 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
775 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
776 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
777 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
778 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
779 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
780 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
781 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
782 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
783 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
784 { X86::SBB32rr, X86::SBB32rm, 0 },
785 { X86::SBB64rr, X86::SBB64rm, 0 },
786 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
787 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
788 { X86::SUB16rr, X86::SUB16rm, 0 },
789 { X86::SUB32rr, X86::SUB32rm, 0 },
790 { X86::SUB64rr, X86::SUB64rm, 0 },
791 { X86::SUB8rr, X86::SUB8rm, 0 },
792 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
793 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
794 { X86::SUBSDrr, X86::SUBSDrm, 0 },
795 { X86::SUBSSrr, X86::SUBSSrm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]796 // FIXME: TEST*rr -> swapped operand of TEST*mr.
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]797 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
798 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
799 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
800 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
801 { X86::XOR16rr, X86::XOR16rm, 0 },
802 { X86::XOR32rr, X86::XOR32rm, 0 },
803 { X86::XOR64rr, X86::XOR64rm, 0 },
804 { X86::XOR8rr, X86::XOR8rm, 0 },
805 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]806 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
807 // AVX 128-bit versions of foldable instructions
808 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
809 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
810 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
811 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
812 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
813 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
814 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
815 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
816 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
817 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
818 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
819 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
Craig Topperb6eb5132012-06-25 06:16:00 +0000[diff] [blame]820 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]821 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, TB_ALIGN_16 },
822 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
823 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
824 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
825 { X86::VADDPDrr, X86::VADDPDrm, TB_ALIGN_16 },
826 { X86::VADDPSrr, X86::VADDPSrm, TB_ALIGN_16 },
827 { X86::VADDSDrr, X86::VADDSDrm, 0 },
828 { X86::VADDSSrr, X86::VADDSSrm, 0 },
829 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, TB_ALIGN_16 },
830 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, TB_ALIGN_16 },
831 { X86::VANDNPDrr, X86::VANDNPDrm, TB_ALIGN_16 },
832 { X86::VANDNPSrr, X86::VANDNPSrm, TB_ALIGN_16 },
833 { X86::VANDPDrr, X86::VANDPDrm, TB_ALIGN_16 },
834 { X86::VANDPSrr, X86::VANDPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]835 { X86::VBLENDPDrri, X86::VBLENDPDrmi, TB_ALIGN_16 },
836 { X86::VBLENDPSrri, X86::VBLENDPSrmi, TB_ALIGN_16 },
837 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, TB_ALIGN_16 },
838 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]839 { X86::VCMPPDrri, X86::VCMPPDrmi, TB_ALIGN_16 },
840 { X86::VCMPPSrri, X86::VCMPPSrmi, TB_ALIGN_16 },
841 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
842 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
843 { X86::VDIVPDrr, X86::VDIVPDrm, TB_ALIGN_16 },
844 { X86::VDIVPSrr, X86::VDIVPSrm, TB_ALIGN_16 },
845 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
846 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
847 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
848 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
849 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
850 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
851 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
852 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
853 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
854 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
855 { X86::VHADDPDrr, X86::VHADDPDrm, TB_ALIGN_16 },
856 { X86::VHADDPSrr, X86::VHADDPSrm, TB_ALIGN_16 },
857 { X86::VHSUBPDrr, X86::VHSUBPDrm, TB_ALIGN_16 },
858 { X86::VHSUBPSrr, X86::VHSUBPSrm, TB_ALIGN_16 },
859 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
860 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
861 { X86::VMAXPDrr, X86::VMAXPDrm, TB_ALIGN_16 },
862 { X86::VMAXPDrr_Int, X86::VMAXPDrm_Int, TB_ALIGN_16 },
863 { X86::VMAXPSrr, X86::VMAXPSrm, TB_ALIGN_16 },
864 { X86::VMAXPSrr_Int, X86::VMAXPSrm_Int, TB_ALIGN_16 },
865 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
866 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 },
867 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
868 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 },
869 { X86::VMINPDrr, X86::VMINPDrm, TB_ALIGN_16 },
870 { X86::VMINPDrr_Int, X86::VMINPDrm_Int, TB_ALIGN_16 },
871 { X86::VMINPSrr, X86::VMINPSrm, TB_ALIGN_16 },
872 { X86::VMINPSrr_Int, X86::VMINPSrm_Int, TB_ALIGN_16 },
873 { X86::VMINSDrr, X86::VMINSDrm, 0 },
874 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 },
875 { X86::VMINSSrr, X86::VMINSSrm, 0 },
876 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]877 { X86::VMPSADBWrri, X86::VMPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]878 { X86::VMULPDrr, X86::VMULPDrm, TB_ALIGN_16 },
879 { X86::VMULPSrr, X86::VMULPSrm, TB_ALIGN_16 },
880 { X86::VMULSDrr, X86::VMULSDrm, 0 },
881 { X86::VMULSSrr, X86::VMULSSrm, 0 },
882 { X86::VORPDrr, X86::VORPDrm, TB_ALIGN_16 },
883 { X86::VORPSrr, X86::VORPSrm, TB_ALIGN_16 },
884 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, TB_ALIGN_16 },
885 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]886 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]887 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, TB_ALIGN_16 },
888 { X86::VPADDBrr, X86::VPADDBrm, TB_ALIGN_16 },
889 { X86::VPADDDrr, X86::VPADDDrm, TB_ALIGN_16 },
890 { X86::VPADDQrr, X86::VPADDQrm, TB_ALIGN_16 },
891 { X86::VPADDSBrr, X86::VPADDSBrm, TB_ALIGN_16 },
892 { X86::VPADDSWrr, X86::VPADDSWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]893 { X86::VPADDUSBrr, X86::VPADDUSBrm, TB_ALIGN_16 },
894 { X86::VPADDUSWrr, X86::VPADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]895 { X86::VPADDWrr, X86::VPADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]896 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]897 { X86::VPANDNrr, X86::VPANDNrm, TB_ALIGN_16 },
898 { X86::VPANDrr, X86::VPANDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]899 { X86::VPAVGBrr, X86::VPAVGBrm, TB_ALIGN_16 },
900 { X86::VPAVGWrr, X86::VPAVGWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]901 { X86::VPBLENDWrri, X86::VPBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]902 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, TB_ALIGN_16 },
903 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]904 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]905 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, TB_ALIGN_16 },
906 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, TB_ALIGN_16 },
907 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]908 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]909 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]910 { X86::VPHADDDrr, X86::VPHADDDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]911 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]912 { X86::VPHADDWrr, X86::VPHADDWrm, TB_ALIGN_16 },
913 { X86::VPHSUBDrr, X86::VPHSUBDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]914 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]915 { X86::VPHSUBWrr, X86::VPHSUBWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]916 { X86::VPERMILPDrr, X86::VPERMILPDrm, TB_ALIGN_16 },
917 { X86::VPERMILPSrr, X86::VPERMILPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]918 { X86::VPINSRWrri, X86::VPINSRWrmi, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]919 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]920 { X86::VPMADDWDrr, X86::VPMADDWDrm, TB_ALIGN_16 },
921 { X86::VPMAXSWrr, X86::VPMAXSWrm, TB_ALIGN_16 },
922 { X86::VPMAXUBrr, X86::VPMAXUBrm, TB_ALIGN_16 },
923 { X86::VPMINSWrr, X86::VPMINSWrm, TB_ALIGN_16 },
924 { X86::VPMINUBrr, X86::VPMINUBrm, TB_ALIGN_16 },
925 { X86::VPMULDQrr, X86::VPMULDQrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]926 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]927 { X86::VPMULHUWrr, X86::VPMULHUWrm, TB_ALIGN_16 },
928 { X86::VPMULHWrr, X86::VPMULHWrm, TB_ALIGN_16 },
929 { X86::VPMULLDrr, X86::VPMULLDrm, TB_ALIGN_16 },
930 { X86::VPMULLWrr, X86::VPMULLWrm, TB_ALIGN_16 },
931 { X86::VPMULUDQrr, X86::VPMULUDQrm, TB_ALIGN_16 },
932 { X86::VPORrr, X86::VPORrm, TB_ALIGN_16 },
933 { X86::VPSADBWrr, X86::VPSADBWrm, TB_ALIGN_16 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]934 { X86::VPSHUFBrr, X86::VPSHUFBrm, TB_ALIGN_16 },
935 { X86::VPSIGNBrr, X86::VPSIGNBrm, TB_ALIGN_16 },
936 { X86::VPSIGNWrr, X86::VPSIGNWrm, TB_ALIGN_16 },
937 { X86::VPSIGNDrr, X86::VPSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]938 { X86::VPSLLDrr, X86::VPSLLDrm, TB_ALIGN_16 },
939 { X86::VPSLLQrr, X86::VPSLLQrm, TB_ALIGN_16 },
940 { X86::VPSLLWrr, X86::VPSLLWrm, TB_ALIGN_16 },
941 { X86::VPSRADrr, X86::VPSRADrm, TB_ALIGN_16 },
942 { X86::VPSRAWrr, X86::VPSRAWrm, TB_ALIGN_16 },
943 { X86::VPSRLDrr, X86::VPSRLDrm, TB_ALIGN_16 },
944 { X86::VPSRLQrr, X86::VPSRLQrm, TB_ALIGN_16 },
945 { X86::VPSRLWrr, X86::VPSRLWrm, TB_ALIGN_16 },
946 { X86::VPSUBBrr, X86::VPSUBBrm, TB_ALIGN_16 },
947 { X86::VPSUBDrr, X86::VPSUBDrm, TB_ALIGN_16 },
948 { X86::VPSUBSBrr, X86::VPSUBSBrm, TB_ALIGN_16 },
949 { X86::VPSUBSWrr, X86::VPSUBSWrm, TB_ALIGN_16 },
950 { X86::VPSUBWrr, X86::VPSUBWrm, TB_ALIGN_16 },
951 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, TB_ALIGN_16 },
952 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, TB_ALIGN_16 },
953 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, TB_ALIGN_16 },
954 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, TB_ALIGN_16 },
955 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, TB_ALIGN_16 },
956 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, TB_ALIGN_16 },
957 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, TB_ALIGN_16 },
958 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, TB_ALIGN_16 },
959 { X86::VPXORrr, X86::VPXORrm, TB_ALIGN_16 },
960 { X86::VSHUFPDrri, X86::VSHUFPDrmi, TB_ALIGN_16 },
961 { X86::VSHUFPSrri, X86::VSHUFPSrmi, TB_ALIGN_16 },
962 { X86::VSUBPDrr, X86::VSUBPDrm, TB_ALIGN_16 },
963 { X86::VSUBPSrr, X86::VSUBPSrm, TB_ALIGN_16 },
964 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
965 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
966 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, TB_ALIGN_16 },
967 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, TB_ALIGN_16 },
968 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, TB_ALIGN_16 },
969 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, TB_ALIGN_16 },
970 { X86::VXORPDrr, X86::VXORPDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]971 { X86::VXORPSrr, X86::VXORPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]972 // AVX 256-bit foldable instructions
973 { X86::VADDPDYrr, X86::VADDPDYrm, TB_ALIGN_32 },
974 { X86::VADDPSYrr, X86::VADDPSYrm, TB_ALIGN_32 },
975 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, TB_ALIGN_32 },
976 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, TB_ALIGN_32 },
977 { X86::VANDNPDYrr, X86::VANDNPDYrm, TB_ALIGN_32 },
978 { X86::VANDNPSYrr, X86::VANDNPSYrm, TB_ALIGN_32 },
979 { X86::VANDPDYrr, X86::VANDPDYrm, TB_ALIGN_32 },
980 { X86::VANDPSYrr, X86::VANDPSYrm, TB_ALIGN_32 },
981 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, TB_ALIGN_32 },
982 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, TB_ALIGN_32 },
983 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, TB_ALIGN_32 },
984 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, TB_ALIGN_32 },
985 { X86::VCMPPDYrri, X86::VCMPPDYrmi, TB_ALIGN_32 },
986 { X86::VCMPPSYrri, X86::VCMPPSYrmi, TB_ALIGN_32 },
987 { X86::VDIVPDYrr, X86::VDIVPDYrm, TB_ALIGN_32 },
988 { X86::VDIVPSYrr, X86::VDIVPSYrm, TB_ALIGN_32 },
989 { X86::VHADDPDYrr, X86::VHADDPDYrm, TB_ALIGN_32 },
990 { X86::VHADDPSYrr, X86::VHADDPSYrm, TB_ALIGN_32 },
991 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, TB_ALIGN_32 },
992 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, TB_ALIGN_32 },
993 { X86::VINSERTF128rr, X86::VINSERTF128rm, TB_ALIGN_32 },
994 { X86::VMAXPDYrr, X86::VMAXPDYrm, TB_ALIGN_32 },
995 { X86::VMAXPDYrr_Int, X86::VMAXPDYrm_Int, TB_ALIGN_32 },
996 { X86::VMAXPSYrr, X86::VMAXPSYrm, TB_ALIGN_32 },
997 { X86::VMAXPSYrr_Int, X86::VMAXPSYrm_Int, TB_ALIGN_32 },
998 { X86::VMINPDYrr, X86::VMINPDYrm, TB_ALIGN_32 },
999 { X86::VMINPDYrr_Int, X86::VMINPDYrm_Int, TB_ALIGN_32 },
1000 { X86::VMINPSYrr, X86::VMINPSYrm, TB_ALIGN_32 },
1001 { X86::VMINPSYrr_Int, X86::VMINPSYrm_Int, TB_ALIGN_32 },
1002 { X86::VMULPDYrr, X86::VMULPDYrm, TB_ALIGN_32 },
1003 { X86::VMULPSYrr, X86::VMULPSYrm, TB_ALIGN_32 },
1004 { X86::VORPDYrr, X86::VORPDYrm, TB_ALIGN_32 },
1005 { X86::VORPSYrr, X86::VORPSYrm, TB_ALIGN_32 },
1006 { X86::VPERM2F128rr, X86::VPERM2F128rm, TB_ALIGN_32 },
1007 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, TB_ALIGN_32 },
1008 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, TB_ALIGN_32 },
1009 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, TB_ALIGN_32 },
1010 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, TB_ALIGN_32 },
1011 { X86::VSUBPDYrr, X86::VSUBPDYrm, TB_ALIGN_32 },
1012 { X86::VSUBPSYrr, X86::VSUBPSYrm, TB_ALIGN_32 },
1013 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, TB_ALIGN_32 },
1014 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, TB_ALIGN_32 },
1015 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, TB_ALIGN_32 },
1016 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, TB_ALIGN_32 },
1017 { X86::VXORPDYrr, X86::VXORPDYrm, TB_ALIGN_32 },
1018 { X86::VXORPSYrr, X86::VXORPSYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1019 // AVX2 foldable instructions
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1020 { X86::VINSERTI128rr, X86::VINSERTI128rm, TB_ALIGN_16 },
1021 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, TB_ALIGN_32 },
1022 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, TB_ALIGN_32 },
1023 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, TB_ALIGN_32 },
1024 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, TB_ALIGN_32 },
1025 { X86::VPADDBYrr, X86::VPADDBYrm, TB_ALIGN_32 },
1026 { X86::VPADDDYrr, X86::VPADDDYrm, TB_ALIGN_32 },
1027 { X86::VPADDQYrr, X86::VPADDQYrm, TB_ALIGN_32 },
1028 { X86::VPADDSBYrr, X86::VPADDSBYrm, TB_ALIGN_32 },
1029 { X86::VPADDSWYrr, X86::VPADDSWYrm, TB_ALIGN_32 },
1030 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, TB_ALIGN_32 },
1031 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, TB_ALIGN_32 },
1032 { X86::VPADDWYrr, X86::VPADDWYrm, TB_ALIGN_32 },
1033 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, TB_ALIGN_32 },
1034 { X86::VPANDNYrr, X86::VPANDNYrm, TB_ALIGN_32 },
1035 { X86::VPANDYrr, X86::VPANDYrm, TB_ALIGN_32 },
1036 { X86::VPAVGBYrr, X86::VPAVGBYrm, TB_ALIGN_32 },
1037 { X86::VPAVGWYrr, X86::VPAVGWYrm, TB_ALIGN_32 },
1038 { X86::VPBLENDDrri, X86::VPBLENDDrmi, TB_ALIGN_32 },
1039 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, TB_ALIGN_32 },
1040 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, TB_ALIGN_32 },
1041 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, TB_ALIGN_32 },
1042 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, TB_ALIGN_32 },
1043 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, TB_ALIGN_32 },
1044 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, TB_ALIGN_32 },
1045 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, TB_ALIGN_32 },
1046 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, TB_ALIGN_32 },
1047 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, TB_ALIGN_32 },
1048 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, TB_ALIGN_32 },
1049 { X86::VPERM2I128rr, X86::VPERM2I128rm, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]1050 { X86::VPERMDYrr, X86::VPERMDYrm, TB_ALIGN_32 },
Elena Demikhovsky779a72b2012-04-15 11:18:59 +0000[diff] [blame]1051 { X86::VPERMPDYri, X86::VPERMPDYmi, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]1052 { X86::VPERMPSYrr, X86::VPERMPSYrm, TB_ALIGN_32 },
Elena Demikhovsky779a72b2012-04-15 11:18:59 +0000[diff] [blame]1053 { X86::VPERMQYri, X86::VPERMQYmi, TB_ALIGN_32 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]1054 { X86::VPHADDDYrr, X86::VPHADDDYrm, TB_ALIGN_32 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1055 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, TB_ALIGN_32 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]1056 { X86::VPHADDWYrr, X86::VPHADDWYrm, TB_ALIGN_32 },
1057 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, TB_ALIGN_32 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1058 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, TB_ALIGN_32 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]1059 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, TB_ALIGN_32 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1060 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, TB_ALIGN_32 },
1061 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, TB_ALIGN_32 },
1062 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, TB_ALIGN_32 },
1063 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, TB_ALIGN_32 },
1064 { X86::VPMINSWYrr, X86::VPMINSWYrm, TB_ALIGN_32 },
1065 { X86::VPMINUBYrr, X86::VPMINUBYrm, TB_ALIGN_32 },
1066 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, TB_ALIGN_32 },
1067 { X86::VPMULDQYrr, X86::VPMULDQYrm, TB_ALIGN_32 },
1068 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, TB_ALIGN_32 },
1069 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, TB_ALIGN_32 },
1070 { X86::VPMULHWYrr, X86::VPMULHWYrm, TB_ALIGN_32 },
1071 { X86::VPMULLDYrr, X86::VPMULLDYrm, TB_ALIGN_32 },
1072 { X86::VPMULLWYrr, X86::VPMULLWYrm, TB_ALIGN_32 },
1073 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, TB_ALIGN_32 },
1074 { X86::VPORYrr, X86::VPORYrm, TB_ALIGN_32 },
1075 { X86::VPSADBWYrr, X86::VPSADBWYrm, TB_ALIGN_32 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]1076 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, TB_ALIGN_32 },
1077 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, TB_ALIGN_32 },
1078 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, TB_ALIGN_32 },
1079 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1080 { X86::VPSLLDYrr, X86::VPSLLDYrm, TB_ALIGN_16 },
1081 { X86::VPSLLQYrr, X86::VPSLLQYrm, TB_ALIGN_16 },
1082 { X86::VPSLLWYrr, X86::VPSLLWYrm, TB_ALIGN_16 },
1083 { X86::VPSLLVDrr, X86::VPSLLVDrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1084 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1085 { X86::VPSLLVQrr, X86::VPSLLVQrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1086 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1087 { X86::VPSRADYrr, X86::VPSRADYrm, TB_ALIGN_16 },
1088 { X86::VPSRAWYrr, X86::VPSRAWYrm, TB_ALIGN_16 },
1089 { X86::VPSRAVDrr, X86::VPSRAVDrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1090 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1091 { X86::VPSRLDYrr, X86::VPSRLDYrm, TB_ALIGN_16 },
1092 { X86::VPSRLQYrr, X86::VPSRLQYrm, TB_ALIGN_16 },
1093 { X86::VPSRLWYrr, X86::VPSRLWYrm, TB_ALIGN_16 },
1094 { X86::VPSRLVDrr, X86::VPSRLVDrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1095 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1096 { X86::VPSRLVQrr, X86::VPSRLVQrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1097 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, TB_ALIGN_32 },
1098 { X86::VPSUBBYrr, X86::VPSUBBYrm, TB_ALIGN_32 },
1099 { X86::VPSUBDYrr, X86::VPSUBDYrm, TB_ALIGN_32 },
1100 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, TB_ALIGN_32 },
1101 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, TB_ALIGN_32 },
1102 { X86::VPSUBWYrr, X86::VPSUBWYrm, TB_ALIGN_32 },
1103 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, TB_ALIGN_32 },
1104 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1105 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1106 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, TB_ALIGN_32 },
1107 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, TB_ALIGN_32 },
1108 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, TB_ALIGN_32 },
1109 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, TB_ALIGN_32 },
1110 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, TB_ALIGN_32 },
1111 { X86::VPXORYrr, X86::VPXORYrm, TB_ALIGN_32 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1112 // FIXME: add AVX 256-bit foldable instructions
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1113 };
1114
1115 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]1116 unsigned RegOp = OpTbl2[i].RegOp;
1117 unsigned MemOp = OpTbl2[i].MemOp;
1118 unsigned Flags = OpTbl2[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1119 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1120 RegOp, MemOp,
1121 // Index 2, folded load
1122 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1123 }
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1124
1125 static const X86OpTblEntry OpTbl3[] = {
1126 // FMA foldable instructions
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1127 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1128 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1129 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1130 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1131 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1132 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1133 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, 0 },
1134 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1135
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1136 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1137 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1138 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1139 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1140 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1141 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1142 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1143 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1144 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1145 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1146 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1147 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1148
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1149 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1150 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1151 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1152 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1153 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1154 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1155 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, 0 },
1156 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1157
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1158 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1159 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1160 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1161 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1162 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1163 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1164 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1165 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1166 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1167 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1168 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1169 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1170
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1171 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1172 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1173 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1174 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1175 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1176 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1177 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, 0 },
1178 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1179
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1180 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1181 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1182 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1183 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1184 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1185 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1186 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1187 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1188 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1189 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1190 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1191 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1192
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1193 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1194 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1195 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1196 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1197 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1198 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1199 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, 0 },
1200 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, 0 },
Craig Topper2e127b52012-06-01 05:48:39 +0000[diff] [blame]1201
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1202 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1203 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1204 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1205 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1206 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1207 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1208 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1209 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1210 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1211 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1212 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1213 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1214
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1215 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1216 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1217 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1218 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1219 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1220 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1221 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1222 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1223 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1224 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1225 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1226 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1227
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1228 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1229 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1230 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1231 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1232 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1233 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1234 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1235 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1236 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1237 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1238 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1239 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1240 };
1241
1242 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1243 unsigned RegOp = OpTbl3[i].RegOp;
1244 unsigned MemOp = OpTbl3[i].MemOp;
1245 unsigned Flags = OpTbl3[i].Flags;
1246 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1247 RegOp, MemOp,
1248 // Index 3, folded load
1249 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1250 }
1251
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]1252}
1253
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1254void
1255X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1256 MemOp2RegOpTableType &M2RTable,
1257 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1258 if ((Flags & TB_NO_FORWARD) == 0) {
1259 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1260 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1261 }
1262 if ((Flags & TB_NO_REVERSE) == 0) {
1263 assert(!M2RTable.count(MemOp) &&
1264 "Duplicated entries in unfolding maps?");
1265 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1266 }
1267}
1268
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1269bool
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1270X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1271 unsigned &SrcReg, unsigned &DstReg,
1272 unsigned &SubIdx) const {
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1273 switch (MI.getOpcode()) {
1274 default: break;
1275 case X86::MOVSX16rr8:
1276 case X86::MOVZX16rr8:
1277 case X86::MOVSX32rr8:
1278 case X86::MOVZX32rr8:
1279 case X86::MOVSX64rr8:
1280 case X86::MOVZX64rr8:
Evan Chengceb5a4e2010-01-13 08:01:32 +0000[diff] [blame]1281 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1282 // It's not always legal to reference the low 8-bit of the larger
1283 // register in 32-bit mode.
1284 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1285 case X86::MOVSX32rr16:
1286 case X86::MOVZX32rr16:
1287 case X86::MOVSX64rr16:
1288 case X86::MOVZX64rr16:
1289 case X86::MOVSX64rr32:
1290 case X86::MOVZX64rr32: {
1291 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1292 // Be conservative.
1293 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1294 SrcReg = MI.getOperand(1).getReg();
1295 DstReg = MI.getOperand(0).getReg();
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1296 switch (MI.getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame^]1297 default: llvm_unreachable("Unreachable!");
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1298 case X86::MOVSX16rr8:
1299 case X86::MOVZX16rr8:
1300 case X86::MOVSX32rr8:
1301 case X86::MOVZX32rr8:
1302 case X86::MOVSX64rr8:
1303 case X86::MOVZX64rr8:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1304 SubIdx = X86::sub_8bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1305 break;
1306 case X86::MOVSX32rr16:
1307 case X86::MOVZX32rr16:
1308 case X86::MOVSX64rr16:
1309 case X86::MOVZX64rr16:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1310 SubIdx = X86::sub_16bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1311 break;
1312 case X86::MOVSX64rr32:
1313 case X86::MOVZX64rr32:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1314 SubIdx = X86::sub_32bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1315 break;
1316 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1317 return true;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1318 }
1319 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1320 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1321}
1322
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1323/// isFrameOperand - Return true and the FrameIndex if the specified
1324/// operand and follow operands form a reference to the stack frame.
1325bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1326 int &FrameIndex) const {
1327 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1328 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1329 MI->getOperand(Op+1).getImm() == 1 &&
1330 MI->getOperand(Op+2).getReg() == 0 &&
1331 MI->getOperand(Op+3).getImm() == 0) {
1332 FrameIndex = MI->getOperand(Op).getIndex();
1333 return true;
1334 }
1335 return false;
1336}
1337
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1338static bool isFrameLoadOpcode(int Opcode) {
1339 switch (Opcode) {
David Blaikie46a9f012012-01-20 21:51:11 +0000[diff] [blame]1340 default:
1341 return false;
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1342 case X86::MOV8rm:
1343 case X86::MOV16rm:
1344 case X86::MOV32rm:
Evan Cheng11b0a5d2006-09-08 06:48:29 +0000[diff] [blame]1345 case X86::MOV64rm:
Dale Johannesen3d7008c2007-07-04 21:07:47 +0000[diff] [blame]1346 case X86::LD_Fp64m:
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1347 case X86::MOVSSrm:
1348 case X86::MOVSDrm:
Chris Lattnerbfc2c682006-04-18 16:44:51 +0000[diff] [blame]1349 case X86::MOVAPSrm:
1350 case X86::MOVAPDrm:
Dan Gohmanbdc0f8b2009-01-09 02:40:34 +0000[diff] [blame]1351 case X86::MOVDQArm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1352 case X86::VMOVSSrm:
1353 case X86::VMOVSDrm:
1354 case X86::VMOVAPSrm:
1355 case X86::VMOVAPDrm:
1356 case X86::VMOVDQArm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1357 case X86::VMOVAPSYrm:
1358 case X86::VMOVAPDYrm:
1359 case X86::VMOVDQAYrm:
Bill Wendlinge7b2a862007-04-03 06:00:37 +0000[diff] [blame]1360 case X86::MMX_MOVD64rm:
1361 case X86::MMX_MOVQ64rm:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1362 return true;
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1363 }
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1364}
1365
1366static bool isFrameStoreOpcode(int Opcode) {
1367 switch (Opcode) {
1368 default: break;
1369 case X86::MOV8mr:
1370 case X86::MOV16mr:
1371 case X86::MOV32mr:
1372 case X86::MOV64mr:
1373 case X86::ST_FpP64m:
1374 case X86::MOVSSmr:
1375 case X86::MOVSDmr:
1376 case X86::MOVAPSmr:
1377 case X86::MOVAPDmr:
1378 case X86::MOVDQAmr:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1379 case X86::VMOVSSmr:
1380 case X86::VMOVSDmr:
1381 case X86::VMOVAPSmr:
1382 case X86::VMOVAPDmr:
1383 case X86::VMOVDQAmr:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1384 case X86::VMOVAPSYmr:
1385 case X86::VMOVAPDYmr:
1386 case X86::VMOVDQAYmr:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1387 case X86::MMX_MOVD64mr:
1388 case X86::MMX_MOVQ64mr:
1389 case X86::MMX_MOVNTQmr:
1390 return true;
1391 }
1392 return false;
1393}
1394
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1395unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1396 int &FrameIndex) const {
1397 if (isFrameLoadOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1398 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1399 return MI->getOperand(0).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1400 return 0;
1401}
1402
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1403unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1404 int &FrameIndex) const {
1405 if (isFrameLoadOpcode(MI->getOpcode())) {
1406 unsigned Reg;
1407 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1408 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1409 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1410 const MachineMemOperand *Dummy;
1411 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1412 }
1413 return 0;
1414}
1415
Dan Gohman0b273252008-11-18 19:49:32 +0000[diff] [blame]1416unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1417 int &FrameIndex) const {
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1418 if (isFrameStoreOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1419 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1420 isFrameOperand(MI, 0, FrameIndex))
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]1421 return MI->getOperand(X86::AddrNumOperands).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1422 return 0;
1423}
1424
1425unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1426 int &FrameIndex) const {
1427 if (isFrameStoreOpcode(MI->getOpcode())) {
1428 unsigned Reg;
1429 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1430 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1431 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1432 const MachineMemOperand *Dummy;
1433 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1434 }
1435 return 0;
1436}
1437
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1438/// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1439/// X86::MOVPC32r.
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1440static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
Jakob Stoklund Olesen3b9a4422012-08-08 00:40:47 +0000[diff] [blame]1441 // Don't waste compile time scanning use-def chains of physregs.
1442 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1443 return false;
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1444 bool isPICBase = false;
1445 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1446 E = MRI.def_end(); I != E; ++I) {
1447 MachineInstr *DefMI = I.getOperand().getParent();
1448 if (DefMI->getOpcode() != X86::MOVPC32r)
1449 return false;
1450 assert(!isPICBase && "More than one PIC base?");
1451 isPICBase = true;
1452 }
1453 return isPICBase;
1454}
Evan Cheng1973a462008-03-31 07:54:19 +0000[diff] [blame]1455
Bill Wendling1e117682008-05-12 20:54:26 +0000[diff] [blame]1456bool
Dan Gohmane919de52009-10-10 00:34:18 +0000[diff] [blame]1457X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1458 AliasAnalysis *AA) const {
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1459 switch (MI->getOpcode()) {
1460 default: break;
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1461 case X86::MOV8rm:
1462 case X86::MOV16rm:
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1463 case X86::MOV32rm:
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1464 case X86::MOV64rm:
1465 case X86::LD_Fp64m:
1466 case X86::MOVSSrm:
1467 case X86::MOVSDrm:
1468 case X86::MOVAPSrm:
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]1469 case X86::MOVUPSrm:
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1470 case X86::MOVAPDrm:
Dan Gohmanbdc0f8b2009-01-09 02:40:34 +0000[diff] [blame]1471 case X86::MOVDQArm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1472 case X86::VMOVSSrm:
1473 case X86::VMOVSDrm:
1474 case X86::VMOVAPSrm:
1475 case X86::VMOVUPSrm:
1476 case X86::VMOVAPDrm:
1477 case X86::VMOVDQArm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1478 case X86::VMOVAPSYrm:
1479 case X86::VMOVUPSYrm:
1480 case X86::VMOVAPDYrm:
1481 case X86::VMOVDQAYrm:
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1482 case X86::MMX_MOVD64rm:
Evan Cheng5392cc92009-11-17 09:51:18 +0000[diff] [blame]1483 case X86::MMX_MOVQ64rm:
Bruno Cardoso Lopesaad5e502011-09-03 00:46:45 +0000[diff] [blame]1484 case X86::FsVMOVAPSrm:
1485 case X86::FsVMOVAPDrm:
Evan Cheng5392cc92009-11-17 09:51:18 +0000[diff] [blame]1486 case X86::FsMOVAPSrm:
1487 case X86::FsMOVAPDrm: {
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1488 // Loads from constant pools are trivially rematerializable.
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1489 if (MI->getOperand(1).isReg() &&
1490 MI->getOperand(2).isImm() &&
1491 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
Dan Gohmane919de52009-10-10 00:34:18 +0000[diff] [blame]1492 MI->isInvariantLoad(AA)) {
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1493 unsigned BaseReg = MI->getOperand(1).getReg();
Chris Lattnerfea81da2009-06-27 04:16:01 +0000[diff] [blame]1494 if (BaseReg == 0 || BaseReg == X86::RIP)
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1495 return true;
1496 // Allow re-materialization of PIC load.
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1497 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
Evan Chengb86595f2008-04-01 23:26:12 +0000[diff] [blame]1498 return false;
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1499 const MachineFunction &MF = *MI->getParent()->getParent();
1500 const MachineRegisterInfo &MRI = MF.getRegInfo();
Jakob Stoklund Olesen3b9a4422012-08-08 00:40:47 +0000[diff] [blame]1501 return regIsPICBase(BaseReg, MRI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1502 }
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1503 return false;
Evan Cheng94ba37f2008-02-22 09:25:47 +0000[diff] [blame]1504 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1505
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1506 case X86::LEA32r:
1507 case X86::LEA64r: {
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1508 if (MI->getOperand(2).isImm() &&
1509 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1510 !MI->getOperand(4).isReg()) {
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1511 // lea fi#, lea GV, etc. are all rematerializable.
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1512 if (!MI->getOperand(1).isReg())
Dan Gohman7e922aa2008-09-26 21:30:20 +0000[diff] [blame]1513 return true;
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1514 unsigned BaseReg = MI->getOperand(1).getReg();
1515 if (BaseReg == 0)
1516 return true;
1517 // Allow re-materialization of lea PICBase + x.
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1518 const MachineFunction &MF = *MI->getParent()->getParent();
1519 const MachineRegisterInfo &MRI = MF.getRegInfo();
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1520 return regIsPICBase(BaseReg, MRI);
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1521 }
1522 return false;
1523 }
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1524 }
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1525
Dan Gohmane8c1e422007-06-26 00:48:07 +0000[diff] [blame]1526 // All other instructions marked M_REMATERIALIZABLE are always trivially
1527 // rematerializable.
1528 return true;
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1529}
1530
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1531/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1532/// would clobber the EFLAGS condition register. Note the result may be
1533/// conservative. If it cannot definitely determine the safety after visiting
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1534/// a few instructions in each direction it assumes it's not safe.
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1535static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1536 MachineBasicBlock::iterator I) {
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1537 MachineBasicBlock::iterator E = MBB.end();
1538
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1539 // For compile time consideration, if we are not able to determine the
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1540 // safety after visiting 4 instructions in each direction, we will assume
1541 // it's not safe.
1542 MachineBasicBlock::iterator Iter = I;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1543 for (unsigned i = 0; Iter != E && i < 4; ++i) {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1544 bool SeenDef = false;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1545 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1546 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1547 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1548 SeenDef = true;
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1549 if (!MO.isReg())
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1550 continue;
1551 if (MO.getReg() == X86::EFLAGS) {
1552 if (MO.isUse())
1553 return false;
1554 SeenDef = true;
1555 }
1556 }
1557
1558 if (SeenDef)
1559 // This instruction defines EFLAGS, no need to look any further.
1560 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1561 ++Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1562 // Skip over DBG_VALUE.
1563 while (Iter != E && Iter->isDebugValue())
1564 ++Iter;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1565 }
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1566
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1567 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1568 // live in.
1569 if (Iter == E) {
1570 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1571 SE = MBB.succ_end(); SI != SE; ++SI)
1572 if ((*SI)->isLiveIn(X86::EFLAGS))
1573 return false;
1574 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1575 }
1576
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1577 MachineBasicBlock::iterator B = MBB.begin();
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1578 Iter = I;
1579 for (unsigned i = 0; i < 4; ++i) {
1580 // If we make it to the beginning of the block, it's safe to clobber
1581 // EFLAGS iff EFLAGS is not live-in.
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1582 if (Iter == B)
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1583 return !MBB.isLiveIn(X86::EFLAGS);
1584
1585 --Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1586 // Skip over DBG_VALUE.
1587 while (Iter != B && Iter->isDebugValue())
1588 --Iter;
1589
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1590 bool SawKill = false;
1591 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1592 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1593 // A register mask may clobber EFLAGS, but we should still look for a
1594 // live EFLAGS def.
1595 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1596 SawKill = true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1597 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1598 if (MO.isDef()) return MO.isDead();
1599 if (MO.isKill()) SawKill = true;
1600 }
1601 }
1602
1603 if (SawKill)
1604 // This instruction kills EFLAGS and doesn't redefine it, so
1605 // there's no need to look further.
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1606 return true;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1607 }
1608
1609 // Conservative answer.
1610 return false;
1611}
1612
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1613void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1614 MachineBasicBlock::iterator I,
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1615 unsigned DestReg, unsigned SubIdx,
Evan Cheng6ad7da92009-11-14 02:55:43 +0000[diff] [blame]1616 const MachineInstr *Orig,
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1617 const TargetRegisterInfo &TRI) const {
Dan Gohman90c600d2010-05-07 01:28:10 +0000[diff] [blame]1618 DebugLoc DL = Orig->getDebugLoc();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1619
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1620 // MOV32r0 etc. are implemented with xor which clobbers condition code.
1621 // Re-materialize them as movri instructions to avoid side effects.
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1622 bool Clone = true;
1623 unsigned Opc = Orig->getOpcode();
1624 switch (Opc) {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1625 default: break;
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1626 case X86::MOV8r0:
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]1627 case X86::MOV16r0:
1628 case X86::MOV32r0:
1629 case X86::MOV64r0: {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1630 if (!isSafeToClobberEFLAGS(MBB, I)) {
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1631 switch (Opc) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame^]1632 default: llvm_unreachable("Unreachable!");
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1633 case X86::MOV8r0: Opc = X86::MOV8ri; break;
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]1634 case X86::MOV16r0: Opc = X86::MOV16ri; break;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1635 case X86::MOV32r0: Opc = X86::MOV32ri; break;
Dan Gohman952f6f92010-02-26 16:49:27 +0000[diff] [blame]1636 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1637 }
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1638 Clone = false;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1639 }
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1640 break;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1641 }
1642 }
1643
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1644 if (Clone) {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1645 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1646 MBB.insert(I, MI);
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1647 } else {
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1648 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1649 }
Evan Cheng147cb762008-04-16 23:44:44 +0000[diff] [blame]1650
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1651 MachineInstr *NewMI = prior(I);
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1652 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1653}
1654
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1655/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1656/// is not marked dead.
1657static bool hasLiveCondCodeDef(MachineInstr *MI) {
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1658 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1659 MachineOperand &MO = MI->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1660 if (MO.isReg() && MO.isDef() &&
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1661 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1662 return true;
1663 }
1664 }
1665 return false;
1666}
1667
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1668/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1669/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1670/// to a 32-bit superregister and then truncating back down to a 16-bit
1671/// subregister.
1672MachineInstr *
1673X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1674 MachineFunction::iterator &MFI,
1675 MachineBasicBlock::iterator &MBBI,
1676 LiveVariables *LV) const {
1677 MachineInstr *MI = MBBI;
1678 unsigned Dest = MI->getOperand(0).getReg();
1679 unsigned Src = MI->getOperand(1).getReg();
1680 bool isDead = MI->getOperand(0).isDead();
1681 bool isKill = MI->getOperand(1).isKill();
1682
1683 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1684 ? X86::LEA64_32r : X86::LEA32r;
1685 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1686 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1687 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1688
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1689 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1690 // well be shifting and then extracting the lower 16-bits.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1691 // This has the potential to cause partial register stall. e.g.
Evan Cheng3974c8d2009-12-12 18:55:26 +0000[diff] [blame]1692 // movw (%rbp,%rcx,2), %dx
1693 // leal -65(%rdx), %esi
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1694 // But testing has shown this *does* help performance in 64-bit mode (at
1695 // least on modern x86 machines).
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1696 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1697 MachineInstr *InsMI =
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1698 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1699 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1700 .addReg(Src, getKillRegState(isKill));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1701
1702 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1703 get(Opc), leaOutReg);
1704 switch (MIOpc) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame^]1705 default: llvm_unreachable("Unreachable!");
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1706 case X86::SHL16ri: {
1707 unsigned ShAmt = MI->getOperand(2).getImm();
1708 MIB.addReg(0).addImm(1 << ShAmt)
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1709 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1710 break;
1711 }
1712 case X86::INC16r:
1713 case X86::INC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1714 addRegOffset(MIB, leaInReg, true, 1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1715 break;
1716 case X86::DEC16r:
1717 case X86::DEC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1718 addRegOffset(MIB, leaInReg, true, -1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1719 break;
1720 case X86::ADD16ri:
1721 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]1722 case X86::ADD16ri_DB:
1723 case X86::ADD16ri8_DB:
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1724 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1725 break;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1726 case X86::ADD16rr:
1727 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1728 unsigned Src2 = MI->getOperand(2).getReg();
1729 bool isKill2 = MI->getOperand(2).isKill();
1730 unsigned leaInReg2 = 0;
1731 MachineInstr *InsMI2 = 0;
1732 if (Src == Src2) {
1733 // ADD16rr %reg1028<kill>, %reg1028
1734 // just a single insert_subreg.
1735 addRegReg(MIB, leaInReg, true, leaInReg, false);
1736 } else {
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1737 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1738 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1739 // well be shifting and then extracting the lower 16-bits.
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1740 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1741 InsMI2 =
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1742 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1743 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1744 .addReg(Src2, getKillRegState(isKill2));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1745 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1746 }
1747 if (LV && isKill2 && InsMI2)
1748 LV->replaceKillInstruction(Src2, MI, InsMI2);
1749 break;
1750 }
1751 }
1752
1753 MachineInstr *NewMI = MIB;
1754 MachineInstr *ExtMI =
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1755 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1756 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1757 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1758
1759 if (LV) {
1760 // Update live variables
1761 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1762 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1763 if (isKill)
1764 LV->replaceKillInstruction(Src, MI, InsMI);
1765 if (isDead)
1766 LV->replaceKillInstruction(Dest, MI, ExtMI);
1767 }
1768
1769 return ExtMI;
1770}
1771
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1772/// convertToThreeAddress - This method must be implemented by targets that
1773/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1774/// may be able to convert a two-address instruction into a true
1775/// three-address instruction on demand. This allows the X86 target (for
1776/// example) to convert ADD and SHL instructions into LEA instructions if they
1777/// would require register copies due to two-addressness.
1778///
1779/// This method returns a null pointer if the transformation cannot be
1780/// performed, otherwise it returns the new instruction.
1781///
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1782MachineInstr *
1783X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1784 MachineBasicBlock::iterator &MBBI,
Owen Anderson30cc0282008-07-02 23:41:07 +0000[diff] [blame]1785 LiveVariables *LV) const {
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1786 MachineInstr *MI = MBBI;
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1787 MachineFunction &MF = *MI->getParent()->getParent();
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1788 // All instructions input are two-addr instructions. Get the known operands.
1789 unsigned Dest = MI->getOperand(0).getReg();
1790 unsigned Src = MI->getOperand(1).getReg();
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1791 bool isDead = MI->getOperand(0).isDead();
1792 bool isKill = MI->getOperand(1).isKill();
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1793
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]1794 MachineInstr *NewMI = NULL;
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1795 // 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]1796 // we have better subtarget support, enable the 16-bit LEA generation here.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1797 // 16-bit LEA is also slow on Core2.
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1798 bool DisableLEA16 = true;
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1799 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1800
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1801 unsigned MIOpc = MI->getOpcode();
1802 switch (MIOpc) {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1803 case X86::SHUFPSrri: {
1804 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1805 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1806
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]1807 unsigned B = MI->getOperand(1).getReg();
1808 unsigned C = MI->getOperand(2).getReg();
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1809 if (B != C) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1810 unsigned A = MI->getOperand(0).getReg();
1811 unsigned M = MI->getOperand(3).getImm();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1812 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]1813 .addReg(A, RegState::Define | getDeadRegState(isDead))
1814 .addReg(B, getKillRegState(isKill)).addImm(M);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1815 break;
1816 }
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1817 case X86::SHUFPDrri: {
1818 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1819 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1820
1821 unsigned B = MI->getOperand(1).getReg();
1822 unsigned C = MI->getOperand(2).getReg();
1823 if (B != C) return 0;
1824 unsigned A = MI->getOperand(0).getReg();
1825 unsigned M = MI->getOperand(3).getImm();
1826
1827 // Convert to PSHUFD mask.
1828 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1829
1830 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1831 .addReg(A, RegState::Define | getDeadRegState(isDead))
1832 .addReg(B, getKillRegState(isKill)).addImm(M);
1833 break;
1834 }
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1835 case X86::SHL64ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1836 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1837 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1838 // the flags produced by a shift yet, so this is safe.
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1839 unsigned ShAmt = MI->getOperand(2).getImm();
1840 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1841
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1842 // LEA can't handle RSP.
1843 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1844 !MF.getRegInfo().constrainRegClass(Src, &X86::GR64_NOSPRegClass))
1845 return 0;
1846
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1847 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]1848 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1849 .addReg(0).addImm(1 << ShAmt)
1850 .addReg(Src, getKillRegState(isKill))
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1851 .addImm(0).addReg(0);
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1852 break;
1853 }
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1854 case X86::SHL32ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1855 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1856 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1857 // the flags produced by a shift yet, so this is safe.
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1858 unsigned ShAmt = MI->getOperand(2).getImm();
1859 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1860
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1861 // LEA can't handle ESP.
1862 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1863 !MF.getRegInfo().constrainRegClass(Src, &X86::GR32_NOSPRegClass))
1864 return 0;
1865
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1866 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1867 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]1868 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1869 .addReg(0).addImm(1 << ShAmt)
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1870 .addReg(Src, getKillRegState(isKill)).addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1871 break;
1872 }
1873 case X86::SHL16ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1874 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Evan Cheng189df732007-09-06 00:14:41 +0000[diff] [blame]1875 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1876 // the flags produced by a shift yet, so this is safe.
Evan Cheng189df732007-09-06 00:14:41 +0000[diff] [blame]1877 unsigned ShAmt = MI->getOperand(2).getImm();
1878 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1879
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1880 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1881 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1882 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1883 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1884 .addReg(0).addImm(1 << ShAmt)
1885 .addReg(Src, getKillRegState(isKill))
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1886 .addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1887 break;
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1888 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1889 default: {
1890 // The following opcodes also sets the condition code register(s). Only
1891 // convert them to equivalent lea if the condition code register def's
1892 // are dead!
1893 if (hasLiveCondCodeDef(MI))
1894 return 0;
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1895
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1896 switch (MIOpc) {
1897 default: return 0;
1898 case X86::INC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]1899 case X86::INC32r:
1900 case X86::INC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1901 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]1902 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1903 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]1904 const TargetRegisterClass *RC = MIOpc == X86::INC64r ?
1905 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1906 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1907
1908 // LEA can't handle RSP.
1909 if (TargetRegisterInfo::isVirtualRegister(Src) &&
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]1910 !MF.getRegInfo().constrainRegClass(Src, RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1911 return 0;
1912
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1913 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]1914 .addReg(Dest, RegState::Define |
1915 getDeadRegState(isDead)),
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]1916 Src, isKill, 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1917 break;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1918 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1919 case X86::INC16r:
1920 case X86::INC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1921 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1922 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1923 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1924 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]1925 .addReg(Dest, RegState::Define |
1926 getDeadRegState(isDead)),
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1927 Src, isKill, 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1928 break;
1929 case X86::DEC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]1930 case X86::DEC32r:
1931 case X86::DEC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1932 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]1933 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1934 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]1935 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ?
1936 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1937 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1938 // LEA can't handle RSP.
1939 if (TargetRegisterInfo::isVirtualRegister(Src) &&
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]1940 !MF.getRegInfo().constrainRegClass(Src, RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1941 return 0;
1942
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1943 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]1944 .addReg(Dest, RegState::Define |
1945 getDeadRegState(isDead)),
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]1946 Src, isKill, -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1947 break;
1948 }
1949 case X86::DEC16r:
1950 case X86::DEC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1951 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1952 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1953 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1954 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]1955 .addReg(Dest, RegState::Define |
1956 getDeadRegState(isDead)),
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1957 Src, isKill, -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1958 break;
1959 case X86::ADD64rr:
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1960 case X86::ADD64rr_DB:
1961 case X86::ADD32rr:
1962 case X86::ADD32rr_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1963 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1964 unsigned Opc;
Craig Topper760b1342012-02-22 05:59:10 +0000[diff] [blame]1965 const TargetRegisterClass *RC;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1966 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
1967 Opc = X86::LEA64r;
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]1968 RC = &X86::GR64_NOSPRegClass;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1969 } else {
1970 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]1971 RC = &X86::GR32_NOSPRegClass;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1972 }
1973
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1974
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1975 unsigned Src2 = MI->getOperand(2).getReg();
1976 bool isKill2 = MI->getOperand(2).isKill();
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1977
1978 // LEA can't handle RSP.
1979 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1980 !MF.getRegInfo().constrainRegClass(Src2, RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1981 return 0;
1982
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1983 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]1984 .addReg(Dest, RegState::Define |
1985 getDeadRegState(isDead)),
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1986 Src, isKill, Src2, isKill2);
Nadav Rotem4968e452012-07-16 10:52:25 +0000[diff] [blame]1987
1988 // Preserve undefness of the operands.
1989 bool isUndef = MI->getOperand(1).isUndef();
1990 bool isUndef2 = MI->getOperand(2).isUndef();
1991 NewMI->getOperand(1).setIsUndef(isUndef);
1992 NewMI->getOperand(3).setIsUndef(isUndef2);
1993
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1994 if (LV && isKill2)
1995 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1996 break;
1997 }
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1998 case X86::ADD16rr:
1999 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2000 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2001 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2002 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2003 unsigned Src2 = MI->getOperand(2).getReg();
2004 bool isKill2 = MI->getOperand(2).isKill();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2005 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]2006 .addReg(Dest, RegState::Define |
2007 getDeadRegState(isDead)),
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2008 Src, isKill, Src2, isKill2);
2009 if (LV && isKill2)
2010 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2011 break;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2012 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2013 case X86::ADD64ri32:
2014 case X86::ADD64ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2015 case X86::ADD64ri32_DB:
2016 case X86::ADD64ri8_DB:
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2017 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]2018 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2019 .addReg(Dest, RegState::Define |
2020 getDeadRegState(isDead)),
2021 Src, isKill, MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2022 break;
2023 case X86::ADD32ri:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2024 case X86::ADD32ri8:
2025 case X86::ADD32ri_DB:
2026 case X86::ADD32ri8_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2027 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2028 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]2029 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2030 .addReg(Dest, RegState::Define |
2031 getDeadRegState(isDead)),
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]2032 Src, isKill, MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2033 break;
2034 }
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2035 case X86::ADD16ri:
2036 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2037 case X86::ADD16ri_DB:
2038 case X86::ADD16ri8_DB:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2039 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2040 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2041 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]2042 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2043 .addReg(Dest, RegState::Define |
2044 getDeadRegState(isDead)),
2045 Src, isKill, MI->getOperand(2).getImm());
2046 break;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2047 }
2048 }
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2049 }
2050
Evan Cheng1bc1cae2008-02-07 08:29:53 +0000[diff] [blame]2051 if (!NewMI) return 0;
2052
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2053 if (LV) { // Update live variables
2054 if (isKill)
2055 LV->replaceKillInstruction(Src, MI, NewMI);
2056 if (isDead)
2057 LV->replaceKillInstruction(Dest, MI, NewMI);
2058 }
2059
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2060 MFI->insert(MBBI, NewMI); // Insert the new inst
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]2061 return NewMI;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2062}
2063
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2064/// commuteInstruction - We have a few instructions that must be hacked on to
2065/// commute them.
2066///
Evan Cheng03553bb2008-06-16 07:33:11 +0000[diff] [blame]2067MachineInstr *
2068X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2069 switch (MI->getOpcode()) {
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2070 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2071 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2072 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2073 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2074 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2075 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2076 unsigned Opc;
2077 unsigned Size;
2078 switch (MI->getOpcode()) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2079 default: llvm_unreachable("Unreachable!");
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2080 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2081 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2082 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2083 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2084 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2085 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2086 }
Chris Lattner5c463782007-12-30 20:49:49 +0000[diff] [blame]2087 unsigned Amt = MI->getOperand(3).getImm();
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2088 if (NewMI) {
2089 MachineFunction &MF = *MI->getParent()->getParent();
2090 MI = MF.CloneMachineInstr(MI);
2091 NewMI = false;
Evan Cheng244183e2008-02-13 02:46:49 +0000[diff] [blame]2092 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2093 MI->setDesc(get(Opc));
2094 MI->getOperand(3).setImm(Size-Amt);
2095 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2096 }
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2097 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2098 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2099 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2100 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2101 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2102 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2103 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2104 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2105 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2106 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2107 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2108 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2109 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2110 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2111 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2112 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2113 unsigned Opc;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2114 switch (MI->getOpcode()) {
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2115 default: llvm_unreachable("Unreachable!");
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2116 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2117 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2118 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2119 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2120 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2121 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2122 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2123 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2124 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2125 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2126 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2127 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
Chris Lattner1a1c6002010-10-05 23:00:14 +0000[diff] [blame]2128 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2129 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2130 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2131 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2132 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2133 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2134 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2135 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2136 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2137 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2138 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2139 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2140 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2141 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2142 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2143 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2144 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2145 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2146 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2147 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2148 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2149 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2150 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2151 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2152 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2153 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2154 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2155 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2156 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2157 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2158 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2159 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2160 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2161 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2162 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2163 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2164 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2165 if (NewMI) {
2166 MachineFunction &MF = *MI->getParent()->getParent();
2167 MI = MF.CloneMachineInstr(MI);
2168 NewMI = false;
2169 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]2170 MI->setDesc(get(Opc));
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2171 // Fallthrough intended.
2172 }
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2173 default:
Evan Cheng03553bb2008-06-16 07:33:11 +0000[diff] [blame]2174 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2175 }
2176}
2177
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2178static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2179 switch (BrOpc) {
2180 default: return X86::COND_INVALID;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2181 case X86::JE_4: return X86::COND_E;
2182 case X86::JNE_4: return X86::COND_NE;
2183 case X86::JL_4: return X86::COND_L;
2184 case X86::JLE_4: return X86::COND_LE;
2185 case X86::JG_4: return X86::COND_G;
2186 case X86::JGE_4: return X86::COND_GE;
2187 case X86::JB_4: return X86::COND_B;
2188 case X86::JBE_4: return X86::COND_BE;
2189 case X86::JA_4: return X86::COND_A;
2190 case X86::JAE_4: return X86::COND_AE;
2191 case X86::JS_4: return X86::COND_S;
2192 case X86::JNS_4: return X86::COND_NS;
2193 case X86::JP_4: return X86::COND_P;
2194 case X86::JNP_4: return X86::COND_NP;
2195 case X86::JO_4: return X86::COND_O;
2196 case X86::JNO_4: return X86::COND_NO;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2197 }
2198}
2199
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2200/// getCondFromSETOpc - return condition code of a SET opcode.
2201static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2202 switch (Opc) {
2203 default: return X86::COND_INVALID;
2204 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2205 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2206 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2207 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2208 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2209 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2210 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2211 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2212 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2213 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2214 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2215 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2216 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2217 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2218 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2219 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2220 }
2221}
2222
2223/// getCondFromCmovOpc - return condition code of a CMov opcode.
2224static X86::CondCode getCondFromCMovOpc(unsigned Opc) {
2225 switch (Opc) {
2226 default: return X86::COND_INVALID;
2227 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2228 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2229 return X86::COND_A;
2230 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2231 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2232 return X86::COND_AE;
2233 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2234 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2235 return X86::COND_B;
2236 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2237 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2238 return X86::COND_BE;
2239 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2240 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2241 return X86::COND_E;
2242 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2243 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2244 return X86::COND_G;
2245 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2246 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2247 return X86::COND_GE;
2248 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2249 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2250 return X86::COND_L;
2251 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2252 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2253 return X86::COND_LE;
2254 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2255 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2256 return X86::COND_NE;
2257 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2258 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2259 return X86::COND_NO;
2260 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2261 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2262 return X86::COND_NP;
2263 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2264 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2265 return X86::COND_NS;
2266 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2267 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2268 return X86::COND_O;
2269 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2270 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2271 return X86::COND_P;
2272 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2273 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2274 return X86::COND_S;
2275 }
2276}
2277
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2278unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2279 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2280 default: llvm_unreachable("Illegal condition code!");
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2281 case X86::COND_E: return X86::JE_4;
2282 case X86::COND_NE: return X86::JNE_4;
2283 case X86::COND_L: return X86::JL_4;
2284 case X86::COND_LE: return X86::JLE_4;
2285 case X86::COND_G: return X86::JG_4;
2286 case X86::COND_GE: return X86::JGE_4;
2287 case X86::COND_B: return X86::JB_4;
2288 case X86::COND_BE: return X86::JBE_4;
2289 case X86::COND_A: return X86::JA_4;
2290 case X86::COND_AE: return X86::JAE_4;
2291 case X86::COND_S: return X86::JS_4;
2292 case X86::COND_NS: return X86::JNS_4;
2293 case X86::COND_P: return X86::JP_4;
2294 case X86::COND_NP: return X86::JNP_4;
2295 case X86::COND_O: return X86::JO_4;
2296 case X86::COND_NO: return X86::JNO_4;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2297 }
2298}
2299
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2300/// GetOppositeBranchCondition - Return the inverse of the specified condition,
2301/// e.g. turning COND_E to COND_NE.
2302X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2303 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2304 default: llvm_unreachable("Illegal condition code!");
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2305 case X86::COND_E: return X86::COND_NE;
2306 case X86::COND_NE: return X86::COND_E;
2307 case X86::COND_L: return X86::COND_GE;
2308 case X86::COND_LE: return X86::COND_G;
2309 case X86::COND_G: return X86::COND_LE;
2310 case X86::COND_GE: return X86::COND_L;
2311 case X86::COND_B: return X86::COND_AE;
2312 case X86::COND_BE: return X86::COND_A;
2313 case X86::COND_A: return X86::COND_BE;
2314 case X86::COND_AE: return X86::COND_B;
2315 case X86::COND_S: return X86::COND_NS;
2316 case X86::COND_NS: return X86::COND_S;
2317 case X86::COND_P: return X86::COND_NP;
2318 case X86::COND_NP: return X86::COND_P;
2319 case X86::COND_O: return X86::COND_NO;
2320 case X86::COND_NO: return X86::COND_O;
2321 }
2322}
2323
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2324/// getSwappedCondition - assume the flags are set by MI(a,b), return
2325/// the condition code if we modify the instructions such that flags are
2326/// set by MI(b,a).
Benjamin Kramerabbfe692012-07-13 13:25:15 +0000[diff] [blame]2327static X86::CondCode getSwappedCondition(X86::CondCode CC) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2328 switch (CC) {
2329 default: return X86::COND_INVALID;
2330 case X86::COND_E: return X86::COND_E;
2331 case X86::COND_NE: return X86::COND_NE;
2332 case X86::COND_L: return X86::COND_G;
2333 case X86::COND_LE: return X86::COND_GE;
2334 case X86::COND_G: return X86::COND_L;
2335 case X86::COND_GE: return X86::COND_LE;
2336 case X86::COND_B: return X86::COND_A;
2337 case X86::COND_BE: return X86::COND_AE;
2338 case X86::COND_A: return X86::COND_B;
2339 case X86::COND_AE: return X86::COND_BE;
2340 }
2341}
2342
2343/// getSETFromCond - Return a set opcode for the given condition and
2344/// whether it has memory operand.
2345static unsigned getSETFromCond(X86::CondCode CC,
2346 bool HasMemoryOperand) {
2347 static const unsigned Opc[16][2] = {
2348 { X86::SETAr, X86::SETAm },
2349 { X86::SETAEr, X86::SETAEm },
2350 { X86::SETBr, X86::SETBm },
2351 { X86::SETBEr, X86::SETBEm },
2352 { X86::SETEr, X86::SETEm },
2353 { X86::SETGr, X86::SETGm },
2354 { X86::SETGEr, X86::SETGEm },
2355 { X86::SETLr, X86::SETLm },
2356 { X86::SETLEr, X86::SETLEm },
2357 { X86::SETNEr, X86::SETNEm },
2358 { X86::SETNOr, X86::SETNOm },
2359 { X86::SETNPr, X86::SETNPm },
2360 { X86::SETNSr, X86::SETNSm },
2361 { X86::SETOr, X86::SETOm },
2362 { X86::SETPr, X86::SETPm },
2363 { X86::SETSr, X86::SETSm }
2364 };
2365
2366 assert(CC < 16 && "Can only handle standard cond codes");
2367 return Opc[CC][HasMemoryOperand ? 1 : 0];
2368}
2369
2370/// getCMovFromCond - Return a cmov opcode for the given condition,
2371/// register size in bytes, and operand type.
2372static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2373 bool HasMemoryOperand) {
2374 static const unsigned Opc[32][3] = {
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2375 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2376 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2377 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2378 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2379 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2380 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2381 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2382 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2383 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2384 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2385 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2386 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2387 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2388 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2389 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2390 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2391 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2392 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2393 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2394 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2395 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2396 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2397 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2398 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2399 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2400 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2401 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2402 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2403 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2404 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2405 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2406 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2407 };
2408
2409 assert(CC < 16 && "Can only handle standard cond codes");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2410 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2411 switch(RegBytes) {
2412 default: llvm_unreachable("Illegal register size!");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2413 case 2: return Opc[Idx][0];
2414 case 4: return Opc[Idx][1];
2415 case 8: return Opc[Idx][2];
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2416 }
2417}
2418
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2419bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2420 if (!MI->isTerminator()) return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2421
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2422 // Conditional branch is a special case.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2423 if (MI->isBranch() && !MI->isBarrier())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2424 return true;
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2425 if (!MI->isPredicable())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2426 return true;
2427 return !isPredicated(MI);
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2428}
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2429
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2430bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2431 MachineBasicBlock *&TBB,
2432 MachineBasicBlock *&FBB,
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2433 SmallVectorImpl<MachineOperand> &Cond,
2434 bool AllowModify) const {
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2435 // Start from the bottom of the block and work up, examining the
2436 // terminator instructions.
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2437 MachineBasicBlock::iterator I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2438 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2439 while (I != MBB.begin()) {
2440 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2441 if (I->isDebugValue())
2442 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2443
2444 // Working from the bottom, when we see a non-terminator instruction, we're
2445 // done.
Jakob Stoklund Olesenc30b4dd2010-07-16 17:41:44 +0000[diff] [blame]2446 if (!isUnpredicatedTerminator(I))
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2447 break;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2448
2449 // A terminator that isn't a branch can't easily be handled by this
2450 // analysis.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2451 if (!I->isBranch())
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2452 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2453
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2454 // Handle unconditional branches.
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2455 if (I->getOpcode() == X86::JMP_4) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2456 UnCondBrIter = I;
2457
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2458 if (!AllowModify) {
2459 TBB = I->getOperand(0).getMBB();
Evan Cheng2fa28112009-05-08 06:34:09 +0000[diff] [blame]2460 continue;
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2461 }
2462
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2463 // If the block has any instructions after a JMP, delete them.
Chris Lattnera48f44d2009-12-03 00:50:42 +0000[diff] [blame]2464 while (llvm::next(I) != MBB.end())
2465 llvm::next(I)->eraseFromParent();
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2466
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2467 Cond.clear();
2468 FBB = 0;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2469
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2470 // Delete the JMP if it's equivalent to a fall-through.
2471 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2472 TBB = 0;
2473 I->eraseFromParent();
2474 I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2475 UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2476 continue;
2477 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2478
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2479 // TBB is used to indicate the unconditional destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2480 TBB = I->getOperand(0).getMBB();
2481 continue;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2482 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2483
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2484 // Handle conditional branches.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2485 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2486 if (BranchCode == X86::COND_INVALID)
2487 return true; // Can't handle indirect branch.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2488
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2489 // Working from the bottom, handle the first conditional branch.
2490 if (Cond.empty()) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2491 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2492 if (AllowModify && UnCondBrIter != MBB.end() &&
2493 MBB.isLayoutSuccessor(TargetBB)) {
2494 // If we can modify the code and it ends in something like:
2495 //
2496 // jCC L1
2497 // jmp L2
2498 // L1:
2499 // ...
2500 // L2:
2501 //
2502 // Then we can change this to:
2503 //
2504 // jnCC L2
2505 // L1:
2506 // ...
2507 // L2:
2508 //
2509 // Which is a bit more efficient.
2510 // We conditionally jump to the fall-through block.
2511 BranchCode = GetOppositeBranchCondition(BranchCode);
2512 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2513 MachineBasicBlock::iterator OldInst = I;
2514
2515 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2516 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2517 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2518 .addMBB(TargetBB);
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2519
2520 OldInst->eraseFromParent();
2521 UnCondBrIter->eraseFromParent();
2522
2523 // Restart the analysis.
2524 UnCondBrIter = MBB.end();
2525 I = MBB.end();
2526 continue;
2527 }
2528
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2529 FBB = TBB;
2530 TBB = I->getOperand(0).getMBB();
2531 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2532 continue;
2533 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2534
2535 // Handle subsequent conditional branches. Only handle the case where all
2536 // conditional branches branch to the same destination and their condition
2537 // opcodes fit one of the special multi-branch idioms.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2538 assert(Cond.size() == 1);
2539 assert(TBB);
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2540
2541 // Only handle the case where all conditional branches branch to the same
2542 // destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2543 if (TBB != I->getOperand(0).getMBB())
2544 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2545
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2546 // If the conditions are the same, we can leave them alone.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2547 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2548 if (OldBranchCode == BranchCode)
2549 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2550
2551 // If they differ, see if they fit one of the known patterns. Theoretically,
2552 // we could handle more patterns here, but we shouldn't expect to see them
2553 // if instruction selection has done a reasonable job.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2554 if ((OldBranchCode == X86::COND_NP &&
2555 BranchCode == X86::COND_E) ||
2556 (OldBranchCode == X86::COND_E &&
2557 BranchCode == X86::COND_NP))
2558 BranchCode = X86::COND_NP_OR_E;
2559 else if ((OldBranchCode == X86::COND_P &&
2560 BranchCode == X86::COND_NE) ||
2561 (OldBranchCode == X86::COND_NE &&
2562 BranchCode == X86::COND_P))
2563 BranchCode = X86::COND_NE_OR_P;
2564 else
2565 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2566
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2567 // Update the MachineOperand.
2568 Cond[0].setImm(BranchCode);
Chris Lattner74436002006-10-30 22:27:23 +0000[diff] [blame]2569 }
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2570
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2571 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2572}
2573
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2574unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2575 MachineBasicBlock::iterator I = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2576 unsigned Count = 0;
2577
2578 while (I != MBB.begin()) {
2579 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2580 if (I->isDebugValue())
2581 continue;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2582 if (I->getOpcode() != X86::JMP_4 &&
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2583 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2584 break;
2585 // Remove the branch.
2586 I->eraseFromParent();
2587 I = MBB.end();
2588 ++Count;
2589 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2590
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2591 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2592}
2593
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2594unsigned
2595X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2596 MachineBasicBlock *FBB,
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2597 const SmallVectorImpl<MachineOperand> &Cond,
2598 DebugLoc DL) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2599 // Shouldn't be a fall through.
2600 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
Chris Lattner6fca75e2006-10-21 05:34:23 +0000[diff] [blame]2601 assert((Cond.size() == 1 || Cond.size() == 0) &&
2602 "X86 branch conditions have one component!");
2603
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2604 if (Cond.empty()) {
2605 // Unconditional branch?
2606 assert(!FBB && "Unconditional branch with multiple successors!");
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2607 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2608 return 1;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2609 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2610
2611 // Conditional branch.
2612 unsigned Count = 0;
2613 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2614 switch (CC) {
2615 case X86::COND_NP_OR_E:
2616 // Synthesize NP_OR_E with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2617 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2618 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2619 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2620 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2621 break;
2622 case X86::COND_NE_OR_P:
2623 // Synthesize NE_OR_P with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2624 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2625 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2626 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2627 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2628 break;
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2629 default: {
2630 unsigned Opc = GetCondBranchFromCond(CC);
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2631 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2632 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2633 }
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2634 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2635 if (FBB) {
2636 // Two-way Conditional branch. Insert the second branch.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2637 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2638 ++Count;
2639 }
2640 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2641}
2642
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2643bool X86InstrInfo::
2644canInsertSelect(const MachineBasicBlock &MBB,
2645 const SmallVectorImpl<MachineOperand> &Cond,
2646 unsigned TrueReg, unsigned FalseReg,
2647 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2648 // Not all subtargets have cmov instructions.
2649 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2650 return false;
2651 if (Cond.size() != 1)
2652 return false;
2653 // We cannot do the composite conditions, at least not in SSA form.
2654 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2655 return false;
2656
2657 // Check register classes.
2658 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2659 const TargetRegisterClass *RC =
2660 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2661 if (!RC)
2662 return false;
2663
2664 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2665 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2666 X86::GR32RegClass.hasSubClassEq(RC) ||
2667 X86::GR64RegClass.hasSubClassEq(RC)) {
2668 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2669 // Bridge. Probably Ivy Bridge as well.
2670 CondCycles = 2;
2671 TrueCycles = 2;
2672 FalseCycles = 2;
2673 return true;
2674 }
2675
2676 // Can't do vectors.
2677 return false;
2678}
2679
2680void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2681 MachineBasicBlock::iterator I, DebugLoc DL,
2682 unsigned DstReg,
2683 const SmallVectorImpl<MachineOperand> &Cond,
2684 unsigned TrueReg, unsigned FalseReg) const {
2685 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2686 assert(Cond.size() == 1 && "Invalid Cond array");
2687 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2688 MRI.getRegClass(DstReg)->getSize(),
2689 false/*HasMemoryOperand*/);
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2690 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2691}
2692
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2693/// isHReg - Test if the given register is a physical h register.
2694static bool isHReg(unsigned Reg) {
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2695 return X86::GR8_ABCD_HRegClass.contains(Reg);
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2696}
2697
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2698// Try and copy between VR128/VR64 and GR64 registers.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2699static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2700 bool HasAVX) {
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2701 // SrcReg(VR128) -> DestReg(GR64)
2702 // SrcReg(VR64) -> DestReg(GR64)
2703 // SrcReg(GR64) -> DestReg(VR128)
2704 // SrcReg(GR64) -> DestReg(VR64)
2705
2706 if (X86::GR64RegClass.contains(DestReg)) {
2707 if (X86::VR128RegClass.contains(SrcReg)) {
2708 // Copy from a VR128 register to a GR64 register.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2709 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2710 } else if (X86::VR64RegClass.contains(SrcReg)) {
2711 // Copy from a VR64 register to a GR64 register.
2712 return X86::MOVSDto64rr;
2713 }
2714 } else if (X86::GR64RegClass.contains(SrcReg)) {
2715 // Copy from a GR64 register to a VR128 register.
2716 if (X86::VR128RegClass.contains(DestReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2717 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2718 // Copy from a GR64 register to a VR64 register.
2719 else if (X86::VR64RegClass.contains(DestReg))
2720 return X86::MOV64toSDrr;
2721 }
2722
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2723 // SrcReg(FR32) -> DestReg(GR32)
2724 // SrcReg(GR32) -> DestReg(FR32)
2725
2726 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
2727 // Copy from a FR32 register to a GR32 register.
2728 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
2729
2730 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
2731 // Copy from a GR32 register to a FR32 register.
2732 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
2733
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2734 return 0;
2735}
2736
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2737void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2738 MachineBasicBlock::iterator MI, DebugLoc DL,
2739 unsigned DestReg, unsigned SrcReg,
2740 bool KillSrc) const {
2741 // First deal with the normal symmetric copies.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2742 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2743 unsigned Opc = 0;
2744 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2745 Opc = X86::MOV64rr;
2746 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2747 Opc = X86::MOV32rr;
2748 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2749 Opc = X86::MOV16rr;
2750 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2751 // Copying to or from a physical H register on x86-64 requires a NOREX
2752 // move. Otherwise use a normal move.
2753 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2754 TM.getSubtarget<X86Subtarget>().is64Bit()) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2755 Opc = X86::MOV8rr_NOREX;
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2756 // Both operands must be encodable without an REX prefix.
2757 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2758 "8-bit H register can not be copied outside GR8_NOREX");
2759 } else
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2760 Opc = X86::MOV8rr;
2761 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2762 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]2763 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2764 Opc = X86::VMOVAPSYrr;
Jakob Stoklund Olesenec58a432010-07-08 22:30:35 +0000[diff] [blame]2765 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2766 Opc = X86::MMX_MOVQ64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2767 else
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2768 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2769
2770 if (Opc) {
2771 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2772 .addReg(SrcReg, getKillRegState(KillSrc));
2773 return;
2774 }
2775
2776 // Moving EFLAGS to / from another register requires a push and a pop.
2777 if (SrcReg == X86::EFLAGS) {
2778 if (X86::GR64RegClass.contains(DestReg)) {
2779 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2780 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2781 return;
2782 } else if (X86::GR32RegClass.contains(DestReg)) {
2783 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2784 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2785 return;
2786 }
2787 }
2788 if (DestReg == X86::EFLAGS) {
2789 if (X86::GR64RegClass.contains(SrcReg)) {
2790 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2791 .addReg(SrcReg, getKillRegState(KillSrc));
2792 BuildMI(MBB, MI, DL, get(X86::POPF64));
2793 return;
2794 } else if (X86::GR32RegClass.contains(SrcReg)) {
2795 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2796 .addReg(SrcReg, getKillRegState(KillSrc));
2797 BuildMI(MBB, MI, DL, get(X86::POPF32));
2798 return;
2799 }
2800 }
2801
2802 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2803 << " to " << RI.getName(DestReg) << '\n');
2804 llvm_unreachable("Cannot emit physreg copy instruction");
2805}
2806
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2807static unsigned getLoadStoreRegOpcode(unsigned Reg,
2808 const TargetRegisterClass *RC,
2809 bool isStackAligned,
2810 const TargetMachine &TM,
2811 bool load) {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2812 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2813 switch (RC->getSize()) {
Rafael Espindola6635f982010-07-12 03:43:04 +0000[diff] [blame]2814 default:
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2815 llvm_unreachable("Unknown spill size");
2816 case 1:
2817 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2818 if (TM.getSubtarget<X86Subtarget>().is64Bit())
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2819 // Copying to or from a physical H register on x86-64 requires a NOREX
2820 // move. Otherwise use a normal move.
2821 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2822 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2823 return load ? X86::MOV8rm : X86::MOV8mr;
2824 case 2:
2825 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2826 return load ? X86::MOV16rm : X86::MOV16mr;
2827 case 4:
2828 if (X86::GR32RegClass.hasSubClassEq(RC))
2829 return load ? X86::MOV32rm : X86::MOV32mr;
2830 if (X86::FR32RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2831 return load ?
2832 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2833 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2834 if (X86::RFP32RegClass.hasSubClassEq(RC))
2835 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2836 llvm_unreachable("Unknown 4-byte regclass");
2837 case 8:
2838 if (X86::GR64RegClass.hasSubClassEq(RC))
2839 return load ? X86::MOV64rm : X86::MOV64mr;
2840 if (X86::FR64RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2841 return load ?
2842 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
2843 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2844 if (X86::VR64RegClass.hasSubClassEq(RC))
2845 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2846 if (X86::RFP64RegClass.hasSubClassEq(RC))
2847 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2848 llvm_unreachable("Unknown 8-byte regclass");
2849 case 10:
2850 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2851 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2852 case 16: {
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2853 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2854 // If stack is realigned we can use aligned stores.
2855 if (isStackAligned)
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2856 return load ?
2857 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
2858 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2859 else
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2860 return load ?
2861 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
2862 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
2863 }
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]2864 case 32:
2865 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2866 // If stack is realigned we can use aligned stores.
2867 if (isStackAligned)
2868 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2869 else
2870 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2871 }
2872}
2873
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2874static unsigned getStoreRegOpcode(unsigned SrcReg,
2875 const TargetRegisterClass *RC,
2876 bool isStackAligned,
2877 TargetMachine &TM) {
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2878 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2879}
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2880
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2881
2882static unsigned getLoadRegOpcode(unsigned DestReg,
2883 const TargetRegisterClass *RC,
2884 bool isStackAligned,
2885 const TargetMachine &TM) {
2886 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2887}
2888
2889void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
2890 MachineBasicBlock::iterator MI,
2891 unsigned SrcReg, bool isKill, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]2892 const TargetRegisterClass *RC,
2893 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]2894 const MachineFunction &MF = *MBB.getParent();
Jakob Stoklund Olesenc3c05ed2010-07-27 04:16:58 +0000[diff] [blame]2895 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
2896 "Stack slot too small for store");
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2897 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2898 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]2899 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2900 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]2901 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2902 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]2903 .addReg(SrcReg, getKillRegState(isKill));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2904}
2905
2906void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
2907 bool isKill,
2908 SmallVectorImpl<MachineOperand> &Addr,
2909 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]2910 MachineInstr::mmo_iterator MMOBegin,
2911 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2912 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2913 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2914 bool isAligned = MMOBegin != MMOEnd &&
2915 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2916 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]2917 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]2918 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2919 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]2920 MIB.addOperand(Addr[i]);
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]2921 MIB.addReg(SrcReg, getKillRegState(isKill));
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]2922 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2923 NewMIs.push_back(MIB);
2924}
2925
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2926
2927void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]2928 MachineBasicBlock::iterator MI,
2929 unsigned DestReg, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]2930 const TargetRegisterClass *RC,
2931 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]2932 const MachineFunction &MF = *MBB.getParent();
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2933 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2934 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]2935 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2936 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]2937 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2938 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2939}
2940
2941void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2942 SmallVectorImpl<MachineOperand> &Addr,
2943 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]2944 MachineInstr::mmo_iterator MMOBegin,
2945 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2946 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2947 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2948 bool isAligned = MMOBegin != MMOEnd &&
2949 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2950 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]2951 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]2952 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2953 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]2954 MIB.addOperand(Addr[i]);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]2955 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2956 NewMIs.push_back(MIB);
2957}
2958
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]2959bool X86InstrInfo::
2960analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
2961 int &CmpMask, int &CmpValue) const {
2962 switch (MI->getOpcode()) {
2963 default: break;
2964 case X86::CMP64ri32:
2965 case X86::CMP64ri8:
2966 case X86::CMP32ri:
2967 case X86::CMP32ri8:
2968 case X86::CMP16ri:
2969 case X86::CMP16ri8:
2970 case X86::CMP8ri:
2971 SrcReg = MI->getOperand(0).getReg();
2972 SrcReg2 = 0;
2973 CmpMask = ~0;
2974 CmpValue = MI->getOperand(1).getImm();
2975 return true;
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]2976 // A SUB can be used to perform comparison.
2977 case X86::SUB64rm:
2978 case X86::SUB32rm:
2979 case X86::SUB16rm:
2980 case X86::SUB8rm:
2981 SrcReg = MI->getOperand(1).getReg();
2982 SrcReg2 = 0;
2983 CmpMask = ~0;
2984 CmpValue = 0;
2985 return true;
2986 case X86::SUB64rr:
2987 case X86::SUB32rr:
2988 case X86::SUB16rr:
2989 case X86::SUB8rr:
2990 SrcReg = MI->getOperand(1).getReg();
2991 SrcReg2 = MI->getOperand(2).getReg();
2992 CmpMask = ~0;
2993 CmpValue = 0;
2994 return true;
2995 case X86::SUB64ri32:
2996 case X86::SUB64ri8:
2997 case X86::SUB32ri:
2998 case X86::SUB32ri8:
2999 case X86::SUB16ri:
3000 case X86::SUB16ri8:
3001 case X86::SUB8ri:
3002 SrcReg = MI->getOperand(1).getReg();
3003 SrcReg2 = 0;
3004 CmpMask = ~0;
3005 CmpValue = MI->getOperand(2).getImm();
3006 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3007 case X86::CMP64rr:
3008 case X86::CMP32rr:
3009 case X86::CMP16rr:
3010 case X86::CMP8rr:
3011 SrcReg = MI->getOperand(0).getReg();
3012 SrcReg2 = MI->getOperand(1).getReg();
3013 CmpMask = ~0;
3014 CmpValue = 0;
3015 return true;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3016 case X86::TEST8rr:
3017 case X86::TEST16rr:
3018 case X86::TEST32rr:
3019 case X86::TEST64rr:
3020 SrcReg = MI->getOperand(0).getReg();
3021 if (MI->getOperand(1).getReg() != SrcReg) return false;
3022 // Compare against zero.
3023 SrcReg2 = 0;
3024 CmpMask = ~0;
3025 CmpValue = 0;
3026 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3027 }
3028 return false;
3029}
3030
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3031/// isRedundantFlagInstr - check whether the first instruction, whose only
3032/// purpose is to update flags, can be made redundant.
3033/// CMPrr can be made redundant by SUBrr if the operands are the same.
3034/// This function can be extended later on.
3035/// SrcReg, SrcRegs: register operands for FlagI.
3036/// ImmValue: immediate for FlagI if it takes an immediate.
3037inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3038 unsigned SrcReg2, int ImmValue,
3039 MachineInstr *OI) {
3040 if (((FlagI->getOpcode() == X86::CMP64rr &&
3041 OI->getOpcode() == X86::SUB64rr) ||
3042 (FlagI->getOpcode() == X86::CMP32rr &&
3043 OI->getOpcode() == X86::SUB32rr)||
3044 (FlagI->getOpcode() == X86::CMP16rr &&
3045 OI->getOpcode() == X86::SUB16rr)||
3046 (FlagI->getOpcode() == X86::CMP8rr &&
3047 OI->getOpcode() == X86::SUB8rr)) &&
3048 ((OI->getOperand(1).getReg() == SrcReg &&
3049 OI->getOperand(2).getReg() == SrcReg2) ||
3050 (OI->getOperand(1).getReg() == SrcReg2 &&
3051 OI->getOperand(2).getReg() == SrcReg)))
3052 return true;
3053
3054 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3055 OI->getOpcode() == X86::SUB64ri32) ||
3056 (FlagI->getOpcode() == X86::CMP64ri8 &&
3057 OI->getOpcode() == X86::SUB64ri8) ||
3058 (FlagI->getOpcode() == X86::CMP32ri &&
3059 OI->getOpcode() == X86::SUB32ri) ||
3060 (FlagI->getOpcode() == X86::CMP32ri8 &&
3061 OI->getOpcode() == X86::SUB32ri8) ||
3062 (FlagI->getOpcode() == X86::CMP16ri &&
3063 OI->getOpcode() == X86::SUB16ri) ||
3064 (FlagI->getOpcode() == X86::CMP16ri8 &&
3065 OI->getOpcode() == X86::SUB16ri8) ||
3066 (FlagI->getOpcode() == X86::CMP8ri &&
3067 OI->getOpcode() == X86::SUB8ri)) &&
3068 OI->getOperand(1).getReg() == SrcReg &&
3069 OI->getOperand(2).getImm() == ImmValue)
3070 return true;
3071 return false;
3072}
3073
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3074/// isDefConvertible - check whether the definition can be converted
3075/// to remove a comparison against zero.
3076inline static bool isDefConvertible(MachineInstr *MI) {
3077 switch (MI->getOpcode()) {
3078 default: return false;
3079 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3080 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3081 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3082 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3083 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3084 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3085 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3086 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3087 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3088 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3089 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3090 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3091 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3092 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3093 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3094 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3095 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3096 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3097 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3098 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3099 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3100 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3101 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3102 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3103 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3104 return true;
3105 }
3106}
3107
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3108/// optimizeCompareInstr - Check if there exists an earlier instruction that
3109/// operates on the same source operands and sets flags in the same way as
3110/// Compare; remove Compare if possible.
3111bool X86InstrInfo::
3112optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3113 int CmpMask, int CmpValue,
3114 const MachineRegisterInfo *MRI) const {
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3115 // Check whether we can replace SUB with CMP.
3116 unsigned NewOpcode = 0;
3117 switch (CmpInstr->getOpcode()) {
3118 default: break;
3119 case X86::SUB64ri32:
3120 case X86::SUB64ri8:
3121 case X86::SUB32ri:
3122 case X86::SUB32ri8:
3123 case X86::SUB16ri:
3124 case X86::SUB16ri8:
3125 case X86::SUB8ri:
3126 case X86::SUB64rm:
3127 case X86::SUB32rm:
3128 case X86::SUB16rm:
3129 case X86::SUB8rm:
3130 case X86::SUB64rr:
3131 case X86::SUB32rr:
3132 case X86::SUB16rr:
3133 case X86::SUB8rr: {
3134 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3135 return false;
3136 // There is no use of the destination register, we can replace SUB with CMP.
3137 switch (CmpInstr->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame^]3138 default: llvm_unreachable("Unreachable!");
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3139 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3140 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3141 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3142 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3143 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3144 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3145 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3146 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3147 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3148 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3149 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3150 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3151 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3152 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3153 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3154 }
3155 CmpInstr->setDesc(get(NewOpcode));
3156 CmpInstr->RemoveOperand(0);
3157 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3158 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3159 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3160 return false;
3161 }
3162 }
3163
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3164 // Get the unique definition of SrcReg.
3165 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3166 if (!MI) return false;
3167
3168 // CmpInstr is the first instruction of the BB.
3169 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3170
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3171 // If we are comparing against zero, check whether we can use MI to update
3172 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3173 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3174 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3175 !isDefConvertible(MI)))
3176 return false;
3177
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3178 // We are searching for an earlier instruction that can make CmpInstr
3179 // redundant and that instruction will be saved in Sub.
3180 MachineInstr *Sub = NULL;
3181 const TargetRegisterInfo *TRI = &getRegisterInfo();
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3182
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3183 // We iterate backward, starting from the instruction before CmpInstr and
3184 // stop when reaching the definition of a source register or done with the BB.
3185 // RI points to the instruction before CmpInstr.
3186 // If the definition is in this basic block, RE points to the definition;
3187 // otherwise, RE is the rend of the basic block.
3188 MachineBasicBlock::reverse_iterator
3189 RI = MachineBasicBlock::reverse_iterator(I),
3190 RE = CmpInstr->getParent() == MI->getParent() ?
3191 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3192 CmpInstr->getParent()->rend();
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3193 MachineInstr *Movr0Inst = 0;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3194 for (; RI != RE; ++RI) {
3195 MachineInstr *Instr = &*RI;
3196 // Check whether CmpInstr can be made redundant by the current instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3197 if (!IsCmpZero &&
3198 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3199 Sub = Instr;
3200 break;
3201 }
3202
3203 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3204 Instr->readsRegister(X86::EFLAGS, TRI)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3205 // This instruction modifies or uses EFLAGS.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3206
3207 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3208 // They are safe to move up, if the definition to EFLAGS is dead and
3209 // earlier instructions do not read or write EFLAGS.
3210 if (!Movr0Inst && (Instr->getOpcode() == X86::MOV8r0 ||
3211 Instr->getOpcode() == X86::MOV16r0 ||
3212 Instr->getOpcode() == X86::MOV32r0 ||
3213 Instr->getOpcode() == X86::MOV64r0) &&
3214 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3215 Movr0Inst = Instr;
3216 continue;
3217 }
3218
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3219 // We can't remove CmpInstr.
3220 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3221 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3222 }
3223
3224 // Return false if no candidates exist.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3225 if (!IsCmpZero && !Sub)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3226 return false;
3227
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3228 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3229 Sub->getOperand(2).getReg() == SrcReg);
3230
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3231 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3232 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3233 // If we are done with the basic block, we need to check whether EFLAGS is
3234 // live-out.
3235 bool IsSafe = false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3236 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3237 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3238 for (++I; I != E; ++I) {
3239 const MachineInstr &Instr = *I;
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3240 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3241 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3242 // We should check the usage if this instruction uses and updates EFLAGS.
3243 if (!UseEFLAGS && ModifyEFLAGS) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3244 // It is safe to remove CmpInstr if EFLAGS is updated again.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3245 IsSafe = true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3246 break;
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3247 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3248 if (!UseEFLAGS && !ModifyEFLAGS)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3249 continue;
3250
3251 // EFLAGS is used by this instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3252 X86::CondCode OldCC;
3253 bool OpcIsSET = false;
3254 if (IsCmpZero || IsSwapped) {
3255 // We decode the condition code from opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3256 if (Instr.isBranch())
3257 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3258 else {
3259 OldCC = getCondFromSETOpc(Instr.getOpcode());
3260 if (OldCC != X86::COND_INVALID)
3261 OpcIsSET = true;
3262 else
3263 OldCC = getCondFromCMovOpc(Instr.getOpcode());
3264 }
3265 if (OldCC == X86::COND_INVALID) return false;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3266 }
3267 if (IsCmpZero) {
3268 switch (OldCC) {
3269 default: break;
3270 case X86::COND_A: case X86::COND_AE:
3271 case X86::COND_B: case X86::COND_BE:
3272 case X86::COND_G: case X86::COND_GE:
3273 case X86::COND_L: case X86::COND_LE:
3274 case X86::COND_O: case X86::COND_NO:
3275 // CF and OF are used, we can't perform this optimization.
3276 return false;
3277 }
3278 } else if (IsSwapped) {
3279 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3280 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3281 // We swap the condition code and synthesize the new opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3282 X86::CondCode NewCC = getSwappedCondition(OldCC);
3283 if (NewCC == X86::COND_INVALID) return false;
3284
3285 // Synthesize the new opcode.
3286 bool HasMemoryOperand = Instr.hasOneMemOperand();
3287 unsigned NewOpc;
3288 if (Instr.isBranch())
3289 NewOpc = GetCondBranchFromCond(NewCC);
3290 else if(OpcIsSET)
3291 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3292 else {
3293 unsigned DstReg = Instr.getOperand(0).getReg();
3294 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3295 HasMemoryOperand);
3296 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3297
3298 // Push the MachineInstr to OpsToUpdate.
3299 // If it is safe to remove CmpInstr, the condition code of these
3300 // instructions will be modified.
3301 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3302 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3303 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3304 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3305 IsSafe = true;
3306 break;
3307 }
3308 }
3309
3310 // If EFLAGS is not killed nor re-defined, we should check whether it is
3311 // live-out. If it is live-out, do not optimize.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3312 if ((IsCmpZero || IsSwapped) && !IsSafe) {
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3313 MachineBasicBlock *MBB = CmpInstr->getParent();
3314 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3315 SE = MBB->succ_end(); SI != SE; ++SI)
3316 if ((*SI)->isLiveIn(X86::EFLAGS))
3317 return false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3318 }
3319
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3320 // The instruction to be updated is either Sub or MI.
3321 Sub = IsCmpZero ? MI : Sub;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3322 // Move Movr0Inst to the place right before Sub.
3323 if (Movr0Inst) {
3324 Sub->getParent()->remove(Movr0Inst);
3325 Sub->getParent()->insert(MachineBasicBlock::iterator(Sub), Movr0Inst);
3326 }
3327
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3328 // Make sure Sub instruction defines EFLAGS.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3329 assert(Sub->getNumOperands() >= 2 &&
3330 Sub->getOperand(Sub->getNumOperands()-1).isReg() &&
3331 Sub->getOperand(Sub->getNumOperands()-1).getReg() == X86::EFLAGS &&
3332 "EFLAGS should be the last operand of SUB, ADD, OR, XOR, AND");
3333 Sub->getOperand(Sub->getNumOperands()-1).setIsDef(true);
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3334 CmpInstr->eraseFromParent();
3335
3336 // Modify the condition code of instructions in OpsToUpdate.
3337 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3338 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3339 return true;
3340}
3341
Manman Ren5759d012012-08-02 00:56:42 +0000[diff] [blame]3342/// optimizeLoadInstr - Try to remove the load by folding it to a register
3343/// operand at the use. We fold the load instructions if load defines a virtual
3344/// register, the virtual register is used once in the same BB, and the
3345/// instructions in-between do not load or store, and have no side effects.
3346MachineInstr* X86InstrInfo::
3347optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3348 unsigned &FoldAsLoadDefReg,
3349 MachineInstr *&DefMI) const {
3350 if (FoldAsLoadDefReg == 0)
3351 return 0;
3352 // To be conservative, if there exists another load, clear the load candidate.
3353 if (MI->mayLoad()) {
3354 FoldAsLoadDefReg = 0;
3355 return 0;
3356 }
3357
3358 // Check whether we can move DefMI here.
3359 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3360 assert(DefMI);
3361 bool SawStore = false;
3362 if (!DefMI->isSafeToMove(this, 0, SawStore))
3363 return 0;
3364
3365 // We try to commute MI if possible.
3366 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3367 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3368 // Collect information about virtual register operands of MI.
3369 unsigned SrcOperandId = 0;
3370 bool FoundSrcOperand = false;
3371 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3372 MachineOperand &MO = MI->getOperand(i);
3373 if (!MO.isReg())
3374 continue;
3375 unsigned Reg = MO.getReg();
3376 if (Reg != FoldAsLoadDefReg)
3377 continue;
3378 // Do not fold if we have a subreg use or a def or multiple uses.
3379 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3380 return 0;
3381
3382 SrcOperandId = i;
3383 FoundSrcOperand = true;
3384 }
3385 if (!FoundSrcOperand) return 0;
3386
3387 // Check whether we can fold the def into SrcOperandId.
3388 SmallVector<unsigned, 8> Ops;
3389 Ops.push_back(SrcOperandId);
3390 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3391 if (FoldMI) {
3392 FoldAsLoadDefReg = 0;
3393 return FoldMI;
3394 }
3395
3396 if (Idx == 1) {
3397 // MI was changed but it didn't help, commute it back!
3398 commuteInstruction(MI, false);
3399 return 0;
3400 }
3401
3402 // Check whether we can commute MI and enable folding.
3403 if (MI->isCommutable()) {
3404 MachineInstr *NewMI = commuteInstruction(MI, false);
3405 // Unable to commute.
3406 if (!NewMI) return 0;
3407 if (NewMI != MI) {
3408 // New instruction. It doesn't need to be kept.
3409 NewMI->eraseFromParent();
3410 return 0;
3411 }
3412 }
3413 }
3414 return 0;
3415}
3416
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3417/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3418/// instruction with two undef reads of the register being defined. This is
3419/// used for mapping:
3420/// %xmm4 = V_SET0
3421/// to:
3422/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3423///
3424static bool Expand2AddrUndef(MachineInstr *MI, const MCInstrDesc &Desc) {
3425 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3426 unsigned Reg = MI->getOperand(0).getReg();
3427 MI->setDesc(Desc);
3428
3429 // MachineInstr::addOperand() will insert explicit operands before any
3430 // implicit operands.
3431 MachineInstrBuilder(MI).addReg(Reg, RegState::Undef)
3432 .addReg(Reg, RegState::Undef);
3433 // But we don't trust that.
3434 assert(MI->getOperand(1).getReg() == Reg &&
3435 MI->getOperand(2).getReg() == Reg && "Misplaced operand");
3436 return true;
3437}
3438
3439bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3440 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3441 switch (MI->getOpcode()) {
3442 case X86::V_SET0:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3443 case X86::FsFLD0SS:
3444 case X86::FsFLD0SD:
Jakob Stoklund Olesen024130892011-11-07 19:15:58 +0000[diff] [blame]3445 return Expand2AddrUndef(MI, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
Jakob Stoklund Olesen729abd32011-10-08 18:28:28 +0000[diff] [blame]3446 case X86::TEST8ri_NOREX:
3447 MI->setDesc(get(X86::TEST8ri));
3448 return true;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3449 }
3450 return false;
3451}
3452
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3453MachineInstr*
3454X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
Evan Cheng250e9172010-04-29 01:13:30 +0000[diff] [blame]3455 int FrameIx, uint64_t Offset,
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3456 const MDNode *MDPtr,
3457 DebugLoc DL) const {
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3458 X86AddressMode AM;
3459 AM.BaseType = X86AddressMode::FrameIndexBase;
3460 AM.Base.FrameIndex = FrameIx;
3461 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
3462 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
3463 return &*MIB;
3464}
3465
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3466static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3467 const SmallVectorImpl<MachineOperand> &MOs,
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3468 MachineInstr *MI,
3469 const TargetInstrInfo &TII) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3470 // Create the base instruction with the memory operand as the first part.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3471 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3472 MI->getDebugLoc(), true);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3473 MachineInstrBuilder MIB(NewMI);
3474 unsigned NumAddrOps = MOs.size();
3475 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3476 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3477 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3478 addOffset(MIB, 0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3479
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3480 // Loop over the rest of the ri operands, converting them over.
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3481 unsigned NumOps = MI->getDesc().getNumOperands()-2;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3482 for (unsigned i = 0; i != NumOps; ++i) {
3483 MachineOperand &MO = MI->getOperand(i+2);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3484 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3485 }
3486 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3487 MachineOperand &MO = MI->getOperand(i);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3488 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3489 }
3490 return MIB;
3491}
3492
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3493static MachineInstr *FuseInst(MachineFunction &MF,
3494 unsigned Opcode, unsigned OpNo,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3495 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3496 MachineInstr *MI, const TargetInstrInfo &TII) {
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3497 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3498 MI->getDebugLoc(), true);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3499 MachineInstrBuilder MIB(NewMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3500
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3501 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3502 MachineOperand &MO = MI->getOperand(i);
3503 if (i == OpNo) {
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3504 assert(MO.isReg() && "Expected to fold into reg operand!");
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3505 unsigned NumAddrOps = MOs.size();
3506 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3507 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3508 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3509 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3510 } else {
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3511 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3512 }
3513 }
3514 return MIB;
3515}
3516
3517static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3518 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3519 MachineInstr *MI) {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3520 MachineFunction &MF = *MI->getParent()->getParent();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3521 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3522
3523 unsigned NumAddrOps = MOs.size();
3524 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3525 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3526 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3527 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3528 return MIB.addImm(0);
3529}
3530
3531MachineInstr*
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3532X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3533 MachineInstr *MI, unsigned i,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3534 const SmallVectorImpl<MachineOperand> &MOs,
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3535 unsigned Size, unsigned Align) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3536 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3537 bool isTwoAddrFold = false;
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3538 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3539 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]3540 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3541
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]3542 // FIXME: AsmPrinter doesn't know how to handle
3543 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3544 if (MI->getOpcode() == X86::ADD32ri &&
3545 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3546 return NULL;
3547
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3548 MachineInstr *NewMI = NULL;
3549 // Folding a memory location into the two-address part of a two-address
3550 // instruction is different than folding it other places. It requires
3551 // replacing the *two* registers with the memory location.
3552 if (isTwoAddr && NumOps >= 2 && i < 2 &&
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3553 MI->getOperand(0).isReg() &&
3554 MI->getOperand(1).isReg() &&
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3555 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3556 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3557 isTwoAddrFold = true;
3558 } else if (i == 0) { // If operand 0
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]3559 if (MI->getOpcode() == X86::MOV64r0)
3560 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
3561 else if (MI->getOpcode() == X86::MOV32r0)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3562 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]3563 else if (MI->getOpcode() == X86::MOV16r0)
3564 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3565 else if (MI->getOpcode() == X86::MOV8r0)
3566 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]3567 if (NewMI)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3568 return NewMI;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3569
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3570 OpcodeTablePtr = &RegOp2MemOpTable0;
3571 } else if (i == 1) {
3572 OpcodeTablePtr = &RegOp2MemOpTable1;
3573 } else if (i == 2) {
3574 OpcodeTablePtr = &RegOp2MemOpTable2;
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]3575 } else if (i == 3) {
3576 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3577 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3578
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3579 // If table selected...
3580 if (OpcodeTablePtr) {
3581 // Find the Opcode to fuse
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3582 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3583 OpcodeTablePtr->find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3584 if (I != OpcodeTablePtr->end()) {
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3585 unsigned Opcode = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]3586 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3587 if (Align < MinAlign)
3588 return NULL;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3589 bool NarrowToMOV32rm = false;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3590 if (Size) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]3591 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3592 if (Size < RCSize) {
3593 // Check if it's safe to fold the load. If the size of the object is
3594 // narrower than the load width, then it's not.
3595 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3596 return NULL;
3597 // If this is a 64-bit load, but the spill slot is 32, then we can do
3598 // a 32-bit load which is implicitly zero-extended. This likely is due
3599 // to liveintervalanalysis remat'ing a load from stack slot.
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3600 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3601 return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3602 Opcode = X86::MOV32rm;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3603 NarrowToMOV32rm = true;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3604 }
3605 }
3606
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3607 if (isTwoAddrFold)
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3608 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3609 else
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3610 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3611
3612 if (NarrowToMOV32rm) {
3613 // If this is the special case where we use a MOV32rm to load a 32-bit
3614 // value and zero-extend the top bits. Change the destination register
3615 // to a 32-bit one.
3616 unsigned DstReg = NewMI->getOperand(0).getReg();
3617 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3618 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3619 X86::sub_32bit));
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3620 else
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3621 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3622 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3623 return NewMI;
3624 }
3625 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3626
3627 // No fusion
Jakob Stoklund Olesen51702ec2010-07-09 20:43:09 +0000[diff] [blame]3628 if (PrintFailedFusing && !MI->isCopy())
David Greened589daf2010-01-05 01:29:29 +0000[diff] [blame]3629 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3630 return NULL;
3631}
3632
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3633/// hasPartialRegUpdate - Return true for all instructions that only update
3634/// the first 32 or 64-bits of the destination register and leave the rest
3635/// unmodified. This can be used to avoid folding loads if the instructions
3636/// only update part of the destination register, and the non-updated part is
3637/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3638/// instructions breaks the partial register dependency and it can improve
3639/// performance. e.g.:
3640///
3641/// movss (%rdi), %xmm0
3642/// cvtss2sd %xmm0, %xmm0
3643///
3644/// Instead of
3645/// cvtss2sd (%rdi), %xmm0
3646///
Bruno Cardoso Lopes7b435682011-09-15 23:04:24 +0000[diff] [blame]3647/// FIXME: This should be turned into a TSFlags.
3648///
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3649static bool hasPartialRegUpdate(unsigned Opcode) {
3650 switch (Opcode) {
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]3651 case X86::CVTSI2SSrr:
3652 case X86::CVTSI2SS64rr:
3653 case X86::CVTSI2SDrr:
3654 case X86::CVTSI2SD64rr:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3655 case X86::CVTSD2SSrr:
3656 case X86::Int_CVTSD2SSrr:
3657 case X86::CVTSS2SDrr:
3658 case X86::Int_CVTSS2SDrr:
3659 case X86::RCPSSr:
3660 case X86::RCPSSr_Int:
3661 case X86::ROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3662 case X86::ROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3663 case X86::ROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3664 case X86::ROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3665 case X86::RSQRTSSr:
3666 case X86::RSQRTSSr_Int:
3667 case X86::SQRTSSr:
3668 case X86::SQRTSSr_Int:
3669 // AVX encoded versions
3670 case X86::VCVTSD2SSrr:
3671 case X86::Int_VCVTSD2SSrr:
3672 case X86::VCVTSS2SDrr:
3673 case X86::Int_VCVTSS2SDrr:
3674 case X86::VRCPSSr:
3675 case X86::VROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3676 case X86::VROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3677 case X86::VROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3678 case X86::VROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3679 case X86::VRSQRTSSr:
3680 case X86::VSQRTSSr:
3681 return true;
3682 }
3683
3684 return false;
3685}
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3686
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]3687/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3688/// instructions we would like before a partial register update.
3689unsigned X86InstrInfo::
3690getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3691 const TargetRegisterInfo *TRI) const {
3692 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3693 return 0;
3694
3695 // If MI is marked as reading Reg, the partial register update is wanted.
3696 const MachineOperand &MO = MI->getOperand(0);
3697 unsigned Reg = MO.getReg();
3698 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3699 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3700 return 0;
3701 } else {
3702 if (MI->readsRegister(Reg, TRI))
3703 return 0;
3704 }
3705
3706 // If any of the preceding 16 instructions are reading Reg, insert a
3707 // dependency breaking instruction. The magic number is based on a few
3708 // Nehalem experiments.
3709 return 16;
3710}
3711
3712void X86InstrInfo::
3713breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3714 const TargetRegisterInfo *TRI) const {
3715 unsigned Reg = MI->getOperand(OpNum).getReg();
3716 if (X86::VR128RegClass.contains(Reg)) {
3717 // These instructions are all floating point domain, so xorps is the best
3718 // choice.
3719 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3720 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3721 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3722 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3723 } else if (X86::VR256RegClass.contains(Reg)) {
3724 // Use vxorps to clear the full ymm register.
3725 // It wants to read and write the xmm sub-register.
3726 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3727 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3728 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3729 .addReg(Reg, RegState::ImplicitDefine);
3730 } else
3731 return;
3732 MI->addRegisterKilled(Reg, TRI, true);
3733}
3734
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3735MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3736 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3737 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3738 int FrameIndex) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3739 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3740 if (NoFusing) return NULL;
3741
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3742 // Unless optimizing for size, don't fold to avoid partial
3743 // register update stalls
3744 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3745 hasPartialRegUpdate(MI->getOpcode()))
3746 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]3747
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3748 const MachineFrameInfo *MFI = MF.getFrameInfo();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3749 unsigned Size = MFI->getObjectSize(FrameIndex);
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3750 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3751 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3752 unsigned NewOpc = 0;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3753 unsigned RCSize = 0;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3754 switch (MI->getOpcode()) {
3755 default: return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3756 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
Dan Gohman887dd1c2010-05-18 21:42:03 +0000[diff] [blame]3757 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
3758 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
3759 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3760 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3761 // Check if it's safe to fold the load. If the size of the object is
3762 // narrower than the load width, then it's not.
3763 if (Size < RCSize)
3764 return NULL;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3765 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]3766 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3767 MI->getOperand(1).ChangeToImmediate(0);
3768 } else if (Ops.size() != 1)
3769 return NULL;
3770
3771 SmallVector<MachineOperand,4> MOs;
3772 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3773 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3774}
3775
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3776MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3777 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3778 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3779 MachineInstr *LoadMI) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3780 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3781 if (NoFusing) return NULL;
3782
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3783 // Unless optimizing for size, don't fold to avoid partial
3784 // register update stalls
3785 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3786 hasPartialRegUpdate(MI->getOpcode()))
3787 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]3788
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]3789 // Determine the alignment of the load.
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3790 unsigned Alignment = 0;
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]3791 if (LoadMI->hasOneMemOperand())
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]3792 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3793 else
3794 switch (LoadMI->getOpcode()) {
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]3795 case X86::AVX_SET0PSY:
3796 case X86::AVX_SET0PDY:
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]3797 case X86::AVX2_SETALLONES:
Craig Toppercb7e13d2012-01-13 08:12:35 +0000[diff] [blame]3798 case X86::AVX2_SET0:
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]3799 Alignment = 32;
3800 break;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3801 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3802 case X86::V_SETALLONES:
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]3803 case X86::AVX_SETALLONES:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3804 Alignment = 16;
3805 break;
3806 case X86::FsFLD0SD:
3807 Alignment = 8;
3808 break;
3809 case X86::FsFLD0SS:
3810 Alignment = 4;
3811 break;
3812 default:
Eli Friedman87ef3872011-06-10 01:13:01 +0000[diff] [blame]3813 return 0;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3814 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3815 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3816 unsigned NewOpc = 0;
3817 switch (MI->getOpcode()) {
3818 default: return NULL;
3819 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]3820 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
3821 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
3822 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3823 }
3824 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]3825 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3826 MI->getOperand(1).ChangeToImmediate(0);
3827 } else if (Ops.size() != 1)
3828 return NULL;
3829
Jakob Stoklund Olesen9c473e42010-08-11 23:08:22 +0000[diff] [blame]3830 // Make sure the subregisters match.
3831 // Otherwise we risk changing the size of the load.
3832 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
3833 return NULL;
3834
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]3835 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3836 switch (LoadMI->getOpcode()) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3837 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3838 case X86::V_SETALLONES:
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]3839 case X86::AVX_SET0PSY:
3840 case X86::AVX_SET0PDY:
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]3841 case X86::AVX_SETALLONES:
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]3842 case X86::AVX2_SETALLONES:
Craig Toppercb7e13d2012-01-13 08:12:35 +0000[diff] [blame]3843 case X86::AVX2_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3844 case X86::FsFLD0SD:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3845 case X86::FsFLD0SS: {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3846 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3847 // Create a constant-pool entry and operands to load from it.
3848
Dan Gohman772952f2010-03-09 03:01:40 +0000[diff] [blame]3849 // Medium and large mode can't fold loads this way.
3850 if (TM.getCodeModel() != CodeModel::Small &&
3851 TM.getCodeModel() != CodeModel::Kernel)
3852 return NULL;
3853
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3854 // x86-32 PIC requires a PIC base register for constant pools.
3855 unsigned PICBase = 0;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]3856 if (TM.getRelocationModel() == Reloc::PIC_) {
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]3857 if (TM.getSubtarget<X86Subtarget>().is64Bit())
3858 PICBase = X86::RIP;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]3859 else
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]3860 // FIXME: PICBase = getGlobalBaseReg(&MF);
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]3861 // This doesn't work for several reasons.
3862 // 1. GlobalBaseReg may have been spilled.
3863 // 2. It may not be live at MI.
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3864 return NULL;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]3865 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3866
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3867 // Create a constant-pool entry.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3868 MachineConstantPool &MCP = *MF.getConstantPool();
Chris Lattner229907c2011-07-18 04:54:35 +0000[diff] [blame]3869 Type *Ty;
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]3870 unsigned Opc = LoadMI->getOpcode();
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3871 if (Opc == X86::FsFLD0SS)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3872 Ty = Type::getFloatTy(MF.getFunction()->getContext());
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3873 else if (Opc == X86::FsFLD0SD)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3874 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]3875 else if (Opc == X86::AVX_SET0PSY || Opc == X86::AVX_SET0PDY)
3876 Ty = VectorType::get(Type::getFloatTy(MF.getFunction()->getContext()), 8);
Craig Toppercb7e13d2012-01-13 08:12:35 +0000[diff] [blame]3877 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX2_SET0)
Craig Toppera4c5a472012-01-13 06:12:41 +0000[diff] [blame]3878 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3879 else
3880 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]3881
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]3882 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX_SETALLONES ||
3883 Opc == X86::AVX2_SETALLONES);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]3884 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
3885 Constant::getNullValue(Ty);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3886 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3887
3888 // Create operands to load from the constant pool entry.
3889 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
3890 MOs.push_back(MachineOperand::CreateImm(1));
3891 MOs.push_back(MachineOperand::CreateReg(0, false));
3892 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3893 MOs.push_back(MachineOperand::CreateReg(0, false));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3894 break;
3895 }
3896 default: {
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3897 // Folding a normal load. Just copy the load's address operands.
3898 unsigned NumOps = LoadMI->getDesc().getNumOperands();
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]3899 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3900 MOs.push_back(LoadMI->getOperand(i));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3901 break;
3902 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3903 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3904 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3905}
3906
3907
Dan Gohman33332bc2008-10-16 01:49:15 +0000[diff] [blame]3908bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
3909 const SmallVectorImpl<unsigned> &Ops) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3910 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3911 if (NoFusing) return 0;
3912
3913 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3914 switch (MI->getOpcode()) {
3915 default: return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3916 case X86::TEST8rr:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3917 case X86::TEST16rr:
3918 case X86::TEST32rr:
3919 case X86::TEST64rr:
3920 return true;
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]3921 case X86::ADD32ri:
3922 // FIXME: AsmPrinter doesn't know how to handle
3923 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3924 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3925 return false;
3926 break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3927 }
3928 }
3929
3930 if (Ops.size() != 1)
3931 return false;
3932
3933 unsigned OpNum = Ops[0];
3934 unsigned Opc = MI->getOpcode();
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3935 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3936 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]3937 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3938
3939 // Folding a memory location into the two-address part of a two-address
3940 // instruction is different than folding it other places. It requires
3941 // replacing the *two* registers with the memory location.
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3942 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3943 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3944 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3945 } else if (OpNum == 0) { // If operand 0
3946 switch (Opc) {
Chris Lattner79c136d2009-07-14 20:19:57 +0000[diff] [blame]3947 case X86::MOV8r0:
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]3948 case X86::MOV16r0:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3949 case X86::MOV32r0:
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3950 case X86::MOV64r0: return true;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3951 default: break;
3952 }
3953 OpcodeTablePtr = &RegOp2MemOpTable0;
3954 } else if (OpNum == 1) {
3955 OpcodeTablePtr = &RegOp2MemOpTable1;
3956 } else if (OpNum == 2) {
3957 OpcodeTablePtr = &RegOp2MemOpTable2;
3958 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3959
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]3960 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
3961 return true;
Jakob Stoklund Olesen7a7b55e2010-07-09 20:43:13 +0000[diff] [blame]3962 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3963}
3964
3965bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
3966 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3967 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3968 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3969 MemOp2RegOpTable.find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3970 if (I == MemOp2RegOpTable.end())
3971 return false;
3972 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]3973 unsigned Index = I->second.second & TB_INDEX_MASK;
3974 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3975 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3976 if (UnfoldLoad && !FoldedLoad)
3977 return false;
3978 UnfoldLoad &= FoldedLoad;
3979 if (UnfoldStore && !FoldedStore)
3980 return false;
3981 UnfoldStore &= FoldedStore;
3982
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]3983 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]3984 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]3985 if (!MI->hasOneMemOperand() &&
3986 RC == &X86::VR128RegClass &&
3987 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3988 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
3989 // conservatively assume the address is unaligned. That's bad for
3990 // performance.
3991 return false;
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]3992 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3993 SmallVector<MachineOperand,2> BeforeOps;
3994 SmallVector<MachineOperand,2> AfterOps;
3995 SmallVector<MachineOperand,4> ImpOps;
3996 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3997 MachineOperand &Op = MI->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]3998 if (i >= Index && i < Index + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3999 AddrOps.push_back(Op);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4000 else if (Op.isReg() && Op.isImplicit())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4001 ImpOps.push_back(Op);
4002 else if (i < Index)
4003 BeforeOps.push_back(Op);
4004 else if (i > Index)
4005 AfterOps.push_back(Op);
4006 }
4007
4008 // Emit the load instruction.
4009 if (UnfoldLoad) {
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4010 std::pair<MachineInstr::mmo_iterator,
4011 MachineInstr::mmo_iterator> MMOs =
4012 MF.extractLoadMemRefs(MI->memoperands_begin(),
4013 MI->memoperands_end());
4014 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4015 if (UnfoldStore) {
4016 // Address operands cannot be marked isKill.
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4017 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4018 MachineOperand &MO = NewMIs[0]->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4019 if (MO.isReg())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4020 MO.setIsKill(false);
4021 }
4022 }
4023 }
4024
4025 // Emit the data processing instruction.
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4026 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4027 MachineInstrBuilder MIB(DataMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4028
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4029 if (FoldedStore)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4030 MIB.addReg(Reg, RegState::Define);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4031 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4032 MIB.addOperand(BeforeOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4033 if (FoldedLoad)
4034 MIB.addReg(Reg);
4035 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4036 MIB.addOperand(AfterOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4037 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4038 MachineOperand &MO = ImpOps[i];
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4039 MIB.addReg(MO.getReg(),
4040 getDefRegState(MO.isDef()) |
4041 RegState::Implicit |
4042 getKillRegState(MO.isKill()) |
Evan Cheng0dc101b2009-06-30 08:49:04 +0000[diff] [blame]4043 getDeadRegState(MO.isDead()) |
4044 getUndefRegState(MO.isUndef()));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4045 }
4046 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4047 switch (DataMI->getOpcode()) {
4048 default: break;
4049 case X86::CMP64ri32:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4050 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4051 case X86::CMP32ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4052 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4053 case X86::CMP16ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4054 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4055 case X86::CMP8ri: {
4056 MachineOperand &MO0 = DataMI->getOperand(0);
4057 MachineOperand &MO1 = DataMI->getOperand(1);
4058 if (MO1.getImm() == 0) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame^]4059 unsigned NewOpc;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4060 switch (DataMI->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame^]4061 default: llvm_unreachable("Unreachable!");
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4062 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4063 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4064 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4065 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4066 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4067 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4068 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4069 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4070 DataMI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4071 MO1.ChangeToRegister(MO0.getReg(), false);
4072 }
4073 }
4074 }
4075 NewMIs.push_back(DataMI);
4076
4077 // Emit the store instruction.
4078 if (UnfoldStore) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4079 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4080 std::pair<MachineInstr::mmo_iterator,
4081 MachineInstr::mmo_iterator> MMOs =
4082 MF.extractStoreMemRefs(MI->memoperands_begin(),
4083 MI->memoperands_end());
4084 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4085 }
4086
4087 return true;
4088}
4089
4090bool
4091X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4092 SmallVectorImpl<SDNode*> &NewNodes) const {
Dan Gohman17059682008-07-17 19:10:17 +0000[diff] [blame]4093 if (!N->isMachineOpcode())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4094 return false;
4095
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4096 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4097 MemOp2RegOpTable.find(N->getMachineOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4098 if (I == MemOp2RegOpTable.end())
4099 return false;
4100 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4101 unsigned Index = I->second.second & TB_INDEX_MASK;
4102 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4103 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4104 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4105 MachineFunction &MF = DAG.getMachineFunction();
4106 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4107 unsigned NumDefs = MCID.NumDefs;
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4108 std::vector<SDValue> AddrOps;
4109 std::vector<SDValue> BeforeOps;
4110 std::vector<SDValue> AfterOps;
Dale Johannesen9f3f72f2009-02-06 01:31:28 +0000[diff] [blame]4111 DebugLoc dl = N->getDebugLoc();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4112 unsigned NumOps = N->getNumOperands();
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]4113 for (unsigned i = 0; i != NumOps-1; ++i) {
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4114 SDValue Op = N->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4115 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4116 AddrOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4117 else if (i < Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4118 BeforeOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4119 else if (i > Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4120 AfterOps.push_back(Op);
4121 }
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4122 SDValue Chain = N->getOperand(NumOps-1);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4123 AddrOps.push_back(Chain);
4124
4125 // Emit the load instruction.
4126 SDNode *Load = 0;
4127 if (FoldedLoad) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4128 EVT VT = *RC->vt_begin();
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4129 std::pair<MachineInstr::mmo_iterator,
4130 MachineInstr::mmo_iterator> MMOs =
4131 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4132 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4133 if (!(*MMOs.first) &&
4134 RC == &X86::VR128RegClass &&
4135 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4136 // Do not introduce a slow unaligned load.
4137 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4138 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4139 bool isAligned = (*MMOs.first) &&
4140 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4141 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
4142 VT, MVT::Other, &AddrOps[0], AddrOps.size());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4143 NewNodes.push_back(Load);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4144
4145 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4146 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4147 }
4148
4149 // Emit the data processing instruction.
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4150 std::vector<EVT> VTs;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4151 const TargetRegisterClass *DstRC = 0;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4152 if (MCID.getNumDefs() > 0) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4153 DstRC = getRegClass(MCID, 0, &RI, MF);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4154 VTs.push_back(*DstRC->vt_begin());
4155 }
4156 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4157 EVT VT = N->getValueType(i);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4158 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4159 VTs.push_back(VT);
4160 }
4161 if (Load)
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4162 BeforeOps.push_back(SDValue(Load, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4163 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4164 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0],
4165 BeforeOps.size());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4166 NewNodes.push_back(NewNode);
4167
4168 // Emit the store instruction.
4169 if (FoldedStore) {
4170 AddrOps.pop_back();
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4171 AddrOps.push_back(SDValue(NewNode, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4172 AddrOps.push_back(Chain);
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4173 std::pair<MachineInstr::mmo_iterator,
4174 MachineInstr::mmo_iterator> MMOs =
4175 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4176 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4177 if (!(*MMOs.first) &&
4178 RC == &X86::VR128RegClass &&
4179 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4180 // Do not introduce a slow unaligned store.
4181 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4182 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4183 bool isAligned = (*MMOs.first) &&
4184 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4185 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4186 isAligned, TM),
4187 dl, MVT::Other,
4188 &AddrOps[0], AddrOps.size());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4189 NewNodes.push_back(Store);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4190
4191 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4192 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4193 }
4194
4195 return true;
4196}
4197
4198unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4199 bool UnfoldLoad, bool UnfoldStore,
4200 unsigned *LoadRegIndex) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4201 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4202 MemOp2RegOpTable.find(Opc);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4203 if (I == MemOp2RegOpTable.end())
4204 return 0;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4205 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4206 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4207 if (UnfoldLoad && !FoldedLoad)
4208 return 0;
4209 if (UnfoldStore && !FoldedStore)
4210 return 0;
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4211 if (LoadRegIndex)
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4212 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4213 return I->second.first;
4214}
4215
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4216bool
4217X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4218 int64_t &Offset1, int64_t &Offset2) const {
4219 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4220 return false;
4221 unsigned Opc1 = Load1->getMachineOpcode();
4222 unsigned Opc2 = Load2->getMachineOpcode();
4223 switch (Opc1) {
4224 default: return false;
4225 case X86::MOV8rm:
4226 case X86::MOV16rm:
4227 case X86::MOV32rm:
4228 case X86::MOV64rm:
4229 case X86::LD_Fp32m:
4230 case X86::LD_Fp64m:
4231 case X86::LD_Fp80m:
4232 case X86::MOVSSrm:
4233 case X86::MOVSDrm:
4234 case X86::MMX_MOVD64rm:
4235 case X86::MMX_MOVQ64rm:
4236 case X86::FsMOVAPSrm:
4237 case X86::FsMOVAPDrm:
4238 case X86::MOVAPSrm:
4239 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4240 case X86::MOVAPDrm:
4241 case X86::MOVDQArm:
4242 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4243 // AVX load instructions
4244 case X86::VMOVSSrm:
4245 case X86::VMOVSDrm:
4246 case X86::FsVMOVAPSrm:
4247 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4248 case X86::VMOVAPSrm:
4249 case X86::VMOVUPSrm:
4250 case X86::VMOVAPDrm:
4251 case X86::VMOVDQArm:
4252 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4253 case X86::VMOVAPSYrm:
4254 case X86::VMOVUPSYrm:
4255 case X86::VMOVAPDYrm:
4256 case X86::VMOVDQAYrm:
4257 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4258 break;
4259 }
4260 switch (Opc2) {
4261 default: return false;
4262 case X86::MOV8rm:
4263 case X86::MOV16rm:
4264 case X86::MOV32rm:
4265 case X86::MOV64rm:
4266 case X86::LD_Fp32m:
4267 case X86::LD_Fp64m:
4268 case X86::LD_Fp80m:
4269 case X86::MOVSSrm:
4270 case X86::MOVSDrm:
4271 case X86::MMX_MOVD64rm:
4272 case X86::MMX_MOVQ64rm:
4273 case X86::FsMOVAPSrm:
4274 case X86::FsMOVAPDrm:
4275 case X86::MOVAPSrm:
4276 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4277 case X86::MOVAPDrm:
4278 case X86::MOVDQArm:
4279 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4280 // AVX load instructions
4281 case X86::VMOVSSrm:
4282 case X86::VMOVSDrm:
4283 case X86::FsVMOVAPSrm:
4284 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4285 case X86::VMOVAPSrm:
4286 case X86::VMOVUPSrm:
4287 case X86::VMOVAPDrm:
4288 case X86::VMOVDQArm:
4289 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4290 case X86::VMOVAPSYrm:
4291 case X86::VMOVUPSYrm:
4292 case X86::VMOVAPDYrm:
4293 case X86::VMOVDQAYrm:
4294 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4295 break;
4296 }
4297
4298 // Check if chain operands and base addresses match.
4299 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4300 Load1->getOperand(5) != Load2->getOperand(5))
4301 return false;
4302 // Segment operands should match as well.
4303 if (Load1->getOperand(4) != Load2->getOperand(4))
4304 return false;
4305 // Scale should be 1, Index should be Reg0.
4306 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4307 Load1->getOperand(2) == Load2->getOperand(2)) {
4308 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4309 return false;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4310
4311 // Now let's examine the displacements.
4312 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4313 isa<ConstantSDNode>(Load2->getOperand(3))) {
4314 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4315 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4316 return true;
4317 }
4318 }
4319 return false;
4320}
4321
4322bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4323 int64_t Offset1, int64_t Offset2,
4324 unsigned NumLoads) const {
4325 assert(Offset2 > Offset1);
4326 if ((Offset2 - Offset1) / 8 > 64)
4327 return false;
4328
4329 unsigned Opc1 = Load1->getMachineOpcode();
4330 unsigned Opc2 = Load2->getMachineOpcode();
4331 if (Opc1 != Opc2)
4332 return false; // FIXME: overly conservative?
4333
4334 switch (Opc1) {
4335 default: break;
4336 case X86::LD_Fp32m:
4337 case X86::LD_Fp64m:
4338 case X86::LD_Fp80m:
4339 case X86::MMX_MOVD64rm:
4340 case X86::MMX_MOVQ64rm:
4341 return false;
4342 }
4343
4344 EVT VT = Load1->getValueType(0);
4345 switch (VT.getSimpleVT().SimpleTy) {
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4346 default:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4347 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4348 // have 16 of them to play with.
4349 if (TM.getSubtargetImpl()->is64Bit()) {
4350 if (NumLoads >= 3)
4351 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4352 } else if (NumLoads) {
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4353 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4354 }
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4355 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4356 case MVT::i8:
4357 case MVT::i16:
4358 case MVT::i32:
4359 case MVT::i64:
Evan Cheng16cf9342010-01-22 23:49:11 +0000[diff] [blame]4360 case MVT::f32:
4361 case MVT::f64:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4362 if (NumLoads)
4363 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4364 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4365 }
4366
4367 return true;
4368}
4369
4370
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4371bool X86InstrInfo::
Owen Anderson4f6bf042008-08-14 22:49:33 +0000[diff] [blame]4372ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4373 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4374 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]4375 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4376 return true;
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4377 Cond[0].setImm(GetOppositeBranchCondition(CC));
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4378 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4379}
4380
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4381bool X86InstrInfo::
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4382isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4383 // FIXME: Return false for x87 stack register classes for now. We can't
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4384 // allow any loads of these registers before FpGet_ST0_80.
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4385 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4386 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4387}
4388
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4389/// getGlobalBaseReg - Return a virtual register initialized with the
4390/// the global base register value. Output instructions required to
4391/// initialize the register in the function entry block, if necessary.
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4392///
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4393/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4394///
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4395unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4396 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4397 "X86-64 PIC uses RIP relative addressing");
4398
4399 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4400 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4401 if (GlobalBaseReg != 0)
4402 return GlobalBaseReg;
4403
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4404 // Create the register. The code to initialize it is inserted
4405 // later, by the CGBR pass (below).
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4406 MachineRegisterInfo &RegInfo = MF->getRegInfo();
Jakob Stoklund Olesen38dcd592012-05-20 18:43:00 +0000[diff] [blame]4407 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4408 X86FI->setGlobalBaseReg(GlobalBaseReg);
4409 return GlobalBaseReg;
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4410}
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4411
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4412// These are the replaceable SSE instructions. Some of these have Int variants
4413// that we don't include here. We don't want to replace instructions selected
4414// by intrinsics.
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4415static const uint16_t ReplaceableInstrs[][3] = {
Bruno Cardoso Lopes1401e042010-08-12 02:08:52 +0000[diff] [blame]4416 //PackedSingle PackedDouble PackedInt
Jakob Stoklund Olesendbff4e82010-03-30 22:46:53 +0000[diff] [blame]4417 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4418 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4419 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4420 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4421 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4422 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4423 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4424 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4425 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4426 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4427 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4428 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4429 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4430 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4431 // AVX 128-bit support
4432 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4433 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4434 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4435 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4436 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4437 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4438 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4439 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4440 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4441 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4442 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4443 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4444 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4445 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4446 // AVX 256-bit support
4447 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4448 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4449 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4450 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4451 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4452 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4453};
4454
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4455static const uint16_t ReplaceableInstrsAVX2[][3] = {
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4456 //PackedSingle PackedDouble PackedInt
Craig Topperf87a2be2011-11-09 09:37:21 +0000[diff] [blame]4457 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4458 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4459 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4460 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4461 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4462 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4463 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
Craig Topper12b72de2011-11-29 05:37:58 +0000[diff] [blame]4464 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4465 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4466 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4467 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4468 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4469 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4470 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4471};
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4472
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4473// FIXME: Some shuffle and unpack instructions have equivalents in different
4474// domains, but they require a bit more work than just switching opcodes.
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4475
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4476static const uint16_t *lookup(unsigned opcode, unsigned domain) {
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4477 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4478 if (ReplaceableInstrs[i][domain-1] == opcode)
4479 return ReplaceableInstrs[i];
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4480 return 0;
4481}
4482
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4483static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4484 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4485 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4486 return ReplaceableInstrsAVX2[i];
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4487 return 0;
4488}
4489
4490std::pair<uint16_t, uint16_t>
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]4491X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4492 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4493 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4494 uint16_t validDomains = 0;
4495 if (domain && lookup(MI->getOpcode(), domain))
4496 validDomains = 0xe;
4497 else if (domain && lookupAVX2(MI->getOpcode(), domain))
4498 validDomains = hasAVX2 ? 0xe : 0x6;
4499 return std::make_pair(domain, validDomains);
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4500}
4501
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]4502void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4503 assert(Domain>0 && Domain<4 && "Invalid execution domain");
4504 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4505 assert(dom && "Not an SSE instruction");
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4506 const uint16_t *table = lookup(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]4507 if (!table) { // try the other table
4508 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
4509 "256-bit vector operations only available in AVX2");
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4510 table = lookupAVX2(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]4511 }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4512 assert(table && "Cannot change domain");
4513 MI->setDesc(get(table[Domain-1]));
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4514}
Chris Lattner6a5e7062010-04-26 23:37:21 +0000[diff] [blame]4515
4516/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
4517void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
4518 NopInst.setOpcode(X86::NOOP);
4519}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4520
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]4521bool X86InstrInfo::isHighLatencyDef(int opc) const {
4522 switch (opc) {
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4523 default: return false;
4524 case X86::DIVSDrm:
4525 case X86::DIVSDrm_Int:
4526 case X86::DIVSDrr:
4527 case X86::DIVSDrr_Int:
4528 case X86::DIVSSrm:
4529 case X86::DIVSSrm_Int:
4530 case X86::DIVSSrr:
4531 case X86::DIVSSrr_Int:
4532 case X86::SQRTPDm:
4533 case X86::SQRTPDm_Int:
4534 case X86::SQRTPDr:
4535 case X86::SQRTPDr_Int:
4536 case X86::SQRTPSm:
4537 case X86::SQRTPSm_Int:
4538 case X86::SQRTPSr:
4539 case X86::SQRTPSr_Int:
4540 case X86::SQRTSDm:
4541 case X86::SQRTSDm_Int:
4542 case X86::SQRTSDr:
4543 case X86::SQRTSDr_Int:
4544 case X86::SQRTSSm:
4545 case X86::SQRTSSm_Int:
4546 case X86::SQRTSSr:
4547 case X86::SQRTSSr_Int:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4548 // AVX instructions with high latency
4549 case X86::VDIVSDrm:
4550 case X86::VDIVSDrm_Int:
4551 case X86::VDIVSDrr:
4552 case X86::VDIVSDrr_Int:
4553 case X86::VDIVSSrm:
4554 case X86::VDIVSSrm_Int:
4555 case X86::VDIVSSrr:
4556 case X86::VDIVSSrr_Int:
4557 case X86::VSQRTPDm:
4558 case X86::VSQRTPDm_Int:
4559 case X86::VSQRTPDr:
4560 case X86::VSQRTPDr_Int:
4561 case X86::VSQRTPSm:
4562 case X86::VSQRTPSm_Int:
4563 case X86::VSQRTPSr:
4564 case X86::VSQRTPSr_Int:
4565 case X86::VSQRTSDm:
4566 case X86::VSQRTSDm_Int:
4567 case X86::VSQRTSDr:
4568 case X86::VSQRTSSm:
4569 case X86::VSQRTSSm_Int:
4570 case X86::VSQRTSSr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4571 return true;
4572 }
4573}
4574
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]4575bool X86InstrInfo::
4576hasHighOperandLatency(const InstrItineraryData *ItinData,
4577 const MachineRegisterInfo *MRI,
4578 const MachineInstr *DefMI, unsigned DefIdx,
4579 const MachineInstr *UseMI, unsigned UseIdx) const {
4580 return isHighLatencyDef(DefMI->getOpcode());
4581}
4582
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4583namespace {
4584 /// CGBR - Create Global Base Reg pass. This initializes the PIC
4585 /// global base register for x86-32.
4586 struct CGBR : public MachineFunctionPass {
4587 static char ID;
Owen Andersona7aed182010-08-06 18:33:48 +0000[diff] [blame]4588 CGBR() : MachineFunctionPass(ID) {}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4589
4590 virtual bool runOnMachineFunction(MachineFunction &MF) {
4591 const X86TargetMachine *TM =
4592 static_cast<const X86TargetMachine *>(&MF.getTarget());
4593
4594 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
4595 "X86-64 PIC uses RIP relative addressing");
4596
4597 // Only emit a global base reg in PIC mode.
4598 if (TM->getRelocationModel() != Reloc::PIC_)
4599 return false;
4600
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]4601 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
4602 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4603
4604 // If we didn't need a GlobalBaseReg, don't insert code.
4605 if (GlobalBaseReg == 0)
4606 return false;
4607
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4608 // Insert the set of GlobalBaseReg into the first MBB of the function
4609 MachineBasicBlock &FirstMBB = MF.front();
4610 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
4611 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
4612 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4613 const X86InstrInfo *TII = TM->getInstrInfo();
4614
4615 unsigned PC;
4616 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]4617 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4618 else
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]4619 PC = GlobalBaseReg;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4620
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4621 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4622 // only used in JIT code emission as displacement to pc.
4623 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4624
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4625 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4626 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4627 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4628 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4629 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4630 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4631 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4632 }
4633
4634 return true;
4635 }
4636
4637 virtual const char *getPassName() const {
4638 return "X86 PIC Global Base Reg Initialization";
4639 }
4640
4641 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4642 AU.setPreservesCFG();
4643 MachineFunctionPass::getAnalysisUsage(AU);
4644 }
4645 };
4646}
4647
4648char CGBR::ID = 0;
4649FunctionPass*
4650llvm::createGlobalBaseRegPass() { return new CGBR(); }
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]4651
4652namespace {
4653 struct LDTLSCleanup : public MachineFunctionPass {
4654 static char ID;
4655 LDTLSCleanup() : MachineFunctionPass(ID) {}
4656
4657 virtual bool runOnMachineFunction(MachineFunction &MF) {
4658 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4659 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4660 // No point folding accesses if there isn't at least two.
4661 return false;
4662 }
4663
4664 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4665 return VisitNode(DT->getRootNode(), 0);
4666 }
4667
4668 // Visit the dominator subtree rooted at Node in pre-order.
4669 // If TLSBaseAddrReg is non-null, then use that to replace any
4670 // TLS_base_addr instructions. Otherwise, create the register
4671 // when the first such instruction is seen, and then use it
4672 // as we encounter more instructions.
4673 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4674 MachineBasicBlock *BB = Node->getBlock();
4675 bool Changed = false;
4676
4677 // Traverse the current block.
4678 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4679 ++I) {
4680 switch (I->getOpcode()) {
4681 case X86::TLS_base_addr32:
4682 case X86::TLS_base_addr64:
4683 if (TLSBaseAddrReg)
4684 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4685 else
4686 I = SetRegister(I, &TLSBaseAddrReg);
4687 Changed = true;
4688 break;
4689 default:
4690 break;
4691 }
4692 }
4693
4694 // Visit the children of this block in the dominator tree.
4695 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4696 I != E; ++I) {
4697 Changed |= VisitNode(*I, TLSBaseAddrReg);
4698 }
4699
4700 return Changed;
4701 }
4702
4703 // Replace the TLS_base_addr instruction I with a copy from
4704 // TLSBaseAddrReg, returning the new instruction.
4705 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4706 unsigned TLSBaseAddrReg) {
4707 MachineFunction *MF = I->getParent()->getParent();
4708 const X86TargetMachine *TM =
4709 static_cast<const X86TargetMachine *>(&MF->getTarget());
4710 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4711 const X86InstrInfo *TII = TM->getInstrInfo();
4712
4713 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4714 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4715 TII->get(TargetOpcode::COPY),
4716 is64Bit ? X86::RAX : X86::EAX)
4717 .addReg(TLSBaseAddrReg);
4718
4719 // Erase the TLS_base_addr instruction.
4720 I->eraseFromParent();
4721
4722 return Copy;
4723 }
4724
4725 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4726 // inserting a copy instruction after I. Returns the new instruction.
4727 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4728 MachineFunction *MF = I->getParent()->getParent();
4729 const X86TargetMachine *TM =
4730 static_cast<const X86TargetMachine *>(&MF->getTarget());
4731 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4732 const X86InstrInfo *TII = TM->getInstrInfo();
4733
4734 // Create a virtual register for the TLS base address.
4735 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4736 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4737 ? &X86::GR64RegClass
4738 : &X86::GR32RegClass);
4739
4740 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4741 MachineInstr *Next = I->getNextNode();
4742 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4743 TII->get(TargetOpcode::COPY),
4744 *TLSBaseAddrReg)
4745 .addReg(is64Bit ? X86::RAX : X86::EAX);
4746
4747 return Copy;
4748 }
4749
4750 virtual const char *getPassName() const {
4751 return "Local Dynamic TLS Access Clean-up";
4752 }
4753
4754 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4755 AU.setPreservesCFG();
4756 AU.addRequired<MachineDominatorTree>();
4757 MachineFunctionPass::getAnalysisUsage(AU);
4758 }
4759 };
4760}
4761
4762char LDTLSCleanup::ID = 0;
4763FunctionPass*
4764llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }