Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / 7ea2a52a0c12c1be618908f5f04261f97d517bee / . / llvm / lib / Target / X86 / X86InstrInfo.cpp
blob: 3d41bf587dc230c26b79964c7a3871032075a280 [file] [log] [blame]
Jia Liub22310f2012-02-18 12:03:15 +0000[diff] [blame]1//===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
Misha Brukmanc88330a2005-04-21 23:38:14 +0000[diff] [blame]2//
John Criswell482202a2003-10-20 19:43:21 +0000[diff] [blame]3// The LLVM Compiler Infrastructure
4//
Chris Lattnerf3ebc3f2007-12-29 20:36:04 +0000[diff] [blame]5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
Misha Brukmanc88330a2005-04-21 23:38:14 +0000[diff] [blame]7//
John Criswell482202a2003-10-20 19:43:21 +0000[diff] [blame]8//===----------------------------------------------------------------------===//
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]9//
Chris Lattnerb4d58d72003-01-14 22:00:31 +0000[diff] [blame]10// This file contains the X86 implementation of the TargetInstrInfo class.
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]11//
12//===----------------------------------------------------------------------===//
13
Chris Lattner27d24792002-10-29 21:05:24 +0000[diff] [blame]14#include "X86InstrInfo.h"
Chris Lattner0d808742002-12-03 05:42:53 +0000[diff] [blame]15#include "X86.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]16#include "X86InstrBuilder.h"
Owen Anderson6bb0c522008-01-04 23:57:37 +0000[diff] [blame]17#include "X86MachineFunctionInfo.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]18#include "X86Subtarget.h"
19#include "X86TargetMachine.h"
Owen Andersone2f23a32007-09-07 04:06:50 +0000[diff] [blame]20#include "llvm/ADT/STLExtras.h"
Chandler Carruthed0881b2012-12-03 16:50:05 +0000[diff] [blame]21#include "llvm/CodeGen/LiveVariables.h"
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]22#include "llvm/CodeGen/MachineConstantPool.h"
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]23#include "llvm/CodeGen/MachineDominators.h"
Owen Anderson6bb0c522008-01-04 23:57:37 +0000[diff] [blame]24#include "llvm/CodeGen/MachineFrameInfo.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]25#include "llvm/CodeGen/MachineInstrBuilder.h"
Chris Lattnera10fff52007-12-31 04:13:23 +0000[diff] [blame]26#include "llvm/CodeGen/MachineRegisterInfo.h"
Chandler Carruth9fb823b2013-01-02 11:36:10 +0000[diff] [blame]27#include "llvm/IR/DerivedTypes.h"
28#include "llvm/IR/LLVMContext.h"
Craig Topperb25fda92012-03-17 18:46:09 +0000[diff] [blame]29#include "llvm/MC/MCAsmInfo.h"
Chris Lattner6a5e7062010-04-26 23:37:21 +0000[diff] [blame]30#include "llvm/MC/MCInst.h"
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]31#include "llvm/Support/CommandLine.h"
David Greened589daf2010-01-05 01:29:29 +0000[diff] [blame]32#include "llvm/Support/Debug.h"
Torok Edwin6dd27302009-07-08 18:01:40 +0000[diff] [blame]33#include "llvm/Support/ErrorHandling.h"
34#include "llvm/Support/raw_ostream.h"
Evan Chenge95f3912007-09-25 01:57:46 +0000[diff] [blame]35#include "llvm/Target/TargetOptions.h"
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]36#include <limits>
37
Evan Cheng703a0fb2011-07-01 17:57:27 +0000[diff] [blame]38#define GET_INSTRINFO_CTOR
Evan Cheng1e210d02011-06-28 20:07:07 +0000[diff] [blame]39#include "X86GenInstrInfo.inc"
40
Brian Gaeke960707c2003-11-11 22:41:34 +0000[diff] [blame]41using namespace llvm;
42
Chris Lattnera6f074f2009-08-23 03:41:05 +0000[diff] [blame]43static cl::opt<bool>
44NoFusing("disable-spill-fusing",
45 cl::desc("Disable fusing of spill code into instructions"));
46static cl::opt<bool>
47PrintFailedFusing("print-failed-fuse-candidates",
48 cl::desc("Print instructions that the allocator wants to"
49 " fuse, but the X86 backend currently can't"),
50 cl::Hidden);
51static cl::opt<bool>
52ReMatPICStubLoad("remat-pic-stub-load",
53 cl::desc("Re-materialize load from stub in PIC mode"),
54 cl::init(false), cl::Hidden);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]55
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]56enum {
57 // Select which memory operand is being unfolded.
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]58 // (stored in bits 0 - 3)
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]59 TB_INDEX_0 = 0,
60 TB_INDEX_1 = 1,
61 TB_INDEX_2 = 2,
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]62 TB_INDEX_3 = 3,
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]63 TB_INDEX_MASK = 0xf,
64
65 // Do not insert the reverse map (MemOp -> RegOp) into the table.
66 // This may be needed because there is a many -> one mapping.
67 TB_NO_REVERSE = 1 << 4,
68
69 // Do not insert the forward map (RegOp -> MemOp) into the table.
70 // This is needed for Native Client, which prohibits branch
71 // instructions from using a memory operand.
72 TB_NO_FORWARD = 1 << 5,
73
74 TB_FOLDED_LOAD = 1 << 6,
75 TB_FOLDED_STORE = 1 << 7,
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]76
77 // Minimum alignment required for load/store.
78 // Used for RegOp->MemOp conversion.
79 // (stored in bits 8 - 15)
80 TB_ALIGN_SHIFT = 8,
81 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
82 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
83 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]84 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]85};
86
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]87struct X86OpTblEntry {
88 uint16_t RegOp;
89 uint16_t MemOp;
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]90 uint16_t Flags;
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]91};
92
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]93X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
Evan Cheng703a0fb2011-07-01 17:57:27 +0000[diff] [blame]94 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
95 ? X86::ADJCALLSTACKDOWN64
96 : X86::ADJCALLSTACKDOWN32),
97 (tm.getSubtarget<X86Subtarget>().is64Bit()
98 ? X86::ADJCALLSTACKUP64
99 : X86::ADJCALLSTACKUP32)),
Evan Cheng11b0a5d2006-09-08 06:48:29 +0000[diff] [blame]100 TM(tm), RI(tm, *this) {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]101
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]102 static const X86OpTblEntry OpTbl2Addr[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]103 { X86::ADC32ri, X86::ADC32mi, 0 },
104 { X86::ADC32ri8, X86::ADC32mi8, 0 },
105 { X86::ADC32rr, X86::ADC32mr, 0 },
106 { X86::ADC64ri32, X86::ADC64mi32, 0 },
107 { X86::ADC64ri8, X86::ADC64mi8, 0 },
108 { X86::ADC64rr, X86::ADC64mr, 0 },
109 { X86::ADD16ri, X86::ADD16mi, 0 },
110 { X86::ADD16ri8, X86::ADD16mi8, 0 },
111 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
112 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
113 { X86::ADD16rr, X86::ADD16mr, 0 },
114 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
115 { X86::ADD32ri, X86::ADD32mi, 0 },
116 { X86::ADD32ri8, X86::ADD32mi8, 0 },
117 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
118 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
119 { X86::ADD32rr, X86::ADD32mr, 0 },
120 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
121 { X86::ADD64ri32, X86::ADD64mi32, 0 },
122 { X86::ADD64ri8, X86::ADD64mi8, 0 },
123 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
124 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
125 { X86::ADD64rr, X86::ADD64mr, 0 },
126 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
127 { X86::ADD8ri, X86::ADD8mi, 0 },
128 { X86::ADD8rr, X86::ADD8mr, 0 },
129 { X86::AND16ri, X86::AND16mi, 0 },
130 { X86::AND16ri8, X86::AND16mi8, 0 },
131 { X86::AND16rr, X86::AND16mr, 0 },
132 { X86::AND32ri, X86::AND32mi, 0 },
133 { X86::AND32ri8, X86::AND32mi8, 0 },
134 { X86::AND32rr, X86::AND32mr, 0 },
135 { X86::AND64ri32, X86::AND64mi32, 0 },
136 { X86::AND64ri8, X86::AND64mi8, 0 },
137 { X86::AND64rr, X86::AND64mr, 0 },
138 { X86::AND8ri, X86::AND8mi, 0 },
139 { X86::AND8rr, X86::AND8mr, 0 },
140 { X86::DEC16r, X86::DEC16m, 0 },
141 { X86::DEC32r, X86::DEC32m, 0 },
142 { X86::DEC64_16r, X86::DEC64_16m, 0 },
143 { X86::DEC64_32r, X86::DEC64_32m, 0 },
144 { X86::DEC64r, X86::DEC64m, 0 },
145 { X86::DEC8r, X86::DEC8m, 0 },
146 { X86::INC16r, X86::INC16m, 0 },
147 { X86::INC32r, X86::INC32m, 0 },
148 { X86::INC64_16r, X86::INC64_16m, 0 },
149 { X86::INC64_32r, X86::INC64_32m, 0 },
150 { X86::INC64r, X86::INC64m, 0 },
151 { X86::INC8r, X86::INC8m, 0 },
152 { X86::NEG16r, X86::NEG16m, 0 },
153 { X86::NEG32r, X86::NEG32m, 0 },
154 { X86::NEG64r, X86::NEG64m, 0 },
155 { X86::NEG8r, X86::NEG8m, 0 },
156 { X86::NOT16r, X86::NOT16m, 0 },
157 { X86::NOT32r, X86::NOT32m, 0 },
158 { X86::NOT64r, X86::NOT64m, 0 },
159 { X86::NOT8r, X86::NOT8m, 0 },
160 { X86::OR16ri, X86::OR16mi, 0 },
161 { X86::OR16ri8, X86::OR16mi8, 0 },
162 { X86::OR16rr, X86::OR16mr, 0 },
163 { X86::OR32ri, X86::OR32mi, 0 },
164 { X86::OR32ri8, X86::OR32mi8, 0 },
165 { X86::OR32rr, X86::OR32mr, 0 },
166 { X86::OR64ri32, X86::OR64mi32, 0 },
167 { X86::OR64ri8, X86::OR64mi8, 0 },
168 { X86::OR64rr, X86::OR64mr, 0 },
169 { X86::OR8ri, X86::OR8mi, 0 },
170 { X86::OR8rr, X86::OR8mr, 0 },
171 { X86::ROL16r1, X86::ROL16m1, 0 },
172 { X86::ROL16rCL, X86::ROL16mCL, 0 },
173 { X86::ROL16ri, X86::ROL16mi, 0 },
174 { X86::ROL32r1, X86::ROL32m1, 0 },
175 { X86::ROL32rCL, X86::ROL32mCL, 0 },
176 { X86::ROL32ri, X86::ROL32mi, 0 },
177 { X86::ROL64r1, X86::ROL64m1, 0 },
178 { X86::ROL64rCL, X86::ROL64mCL, 0 },
179 { X86::ROL64ri, X86::ROL64mi, 0 },
180 { X86::ROL8r1, X86::ROL8m1, 0 },
181 { X86::ROL8rCL, X86::ROL8mCL, 0 },
182 { X86::ROL8ri, X86::ROL8mi, 0 },
183 { X86::ROR16r1, X86::ROR16m1, 0 },
184 { X86::ROR16rCL, X86::ROR16mCL, 0 },
185 { X86::ROR16ri, X86::ROR16mi, 0 },
186 { X86::ROR32r1, X86::ROR32m1, 0 },
187 { X86::ROR32rCL, X86::ROR32mCL, 0 },
188 { X86::ROR32ri, X86::ROR32mi, 0 },
189 { X86::ROR64r1, X86::ROR64m1, 0 },
190 { X86::ROR64rCL, X86::ROR64mCL, 0 },
191 { X86::ROR64ri, X86::ROR64mi, 0 },
192 { X86::ROR8r1, X86::ROR8m1, 0 },
193 { X86::ROR8rCL, X86::ROR8mCL, 0 },
194 { X86::ROR8ri, X86::ROR8mi, 0 },
195 { X86::SAR16r1, X86::SAR16m1, 0 },
196 { X86::SAR16rCL, X86::SAR16mCL, 0 },
197 { X86::SAR16ri, X86::SAR16mi, 0 },
198 { X86::SAR32r1, X86::SAR32m1, 0 },
199 { X86::SAR32rCL, X86::SAR32mCL, 0 },
200 { X86::SAR32ri, X86::SAR32mi, 0 },
201 { X86::SAR64r1, X86::SAR64m1, 0 },
202 { X86::SAR64rCL, X86::SAR64mCL, 0 },
203 { X86::SAR64ri, X86::SAR64mi, 0 },
204 { X86::SAR8r1, X86::SAR8m1, 0 },
205 { X86::SAR8rCL, X86::SAR8mCL, 0 },
206 { X86::SAR8ri, X86::SAR8mi, 0 },
207 { X86::SBB32ri, X86::SBB32mi, 0 },
208 { X86::SBB32ri8, X86::SBB32mi8, 0 },
209 { X86::SBB32rr, X86::SBB32mr, 0 },
210 { X86::SBB64ri32, X86::SBB64mi32, 0 },
211 { X86::SBB64ri8, X86::SBB64mi8, 0 },
212 { X86::SBB64rr, X86::SBB64mr, 0 },
213 { X86::SHL16rCL, X86::SHL16mCL, 0 },
214 { X86::SHL16ri, X86::SHL16mi, 0 },
215 { X86::SHL32rCL, X86::SHL32mCL, 0 },
216 { X86::SHL32ri, X86::SHL32mi, 0 },
217 { X86::SHL64rCL, X86::SHL64mCL, 0 },
218 { X86::SHL64ri, X86::SHL64mi, 0 },
219 { X86::SHL8rCL, X86::SHL8mCL, 0 },
220 { X86::SHL8ri, X86::SHL8mi, 0 },
221 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
222 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
223 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
224 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
225 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
226 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
227 { X86::SHR16r1, X86::SHR16m1, 0 },
228 { X86::SHR16rCL, X86::SHR16mCL, 0 },
229 { X86::SHR16ri, X86::SHR16mi, 0 },
230 { X86::SHR32r1, X86::SHR32m1, 0 },
231 { X86::SHR32rCL, X86::SHR32mCL, 0 },
232 { X86::SHR32ri, X86::SHR32mi, 0 },
233 { X86::SHR64r1, X86::SHR64m1, 0 },
234 { X86::SHR64rCL, X86::SHR64mCL, 0 },
235 { X86::SHR64ri, X86::SHR64mi, 0 },
236 { X86::SHR8r1, X86::SHR8m1, 0 },
237 { X86::SHR8rCL, X86::SHR8mCL, 0 },
238 { X86::SHR8ri, X86::SHR8mi, 0 },
239 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
240 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
241 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
242 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
243 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
244 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
245 { X86::SUB16ri, X86::SUB16mi, 0 },
246 { X86::SUB16ri8, X86::SUB16mi8, 0 },
247 { X86::SUB16rr, X86::SUB16mr, 0 },
248 { X86::SUB32ri, X86::SUB32mi, 0 },
249 { X86::SUB32ri8, X86::SUB32mi8, 0 },
250 { X86::SUB32rr, X86::SUB32mr, 0 },
251 { X86::SUB64ri32, X86::SUB64mi32, 0 },
252 { X86::SUB64ri8, X86::SUB64mi8, 0 },
253 { X86::SUB64rr, X86::SUB64mr, 0 },
254 { X86::SUB8ri, X86::SUB8mi, 0 },
255 { X86::SUB8rr, X86::SUB8mr, 0 },
256 { X86::XOR16ri, X86::XOR16mi, 0 },
257 { X86::XOR16ri8, X86::XOR16mi8, 0 },
258 { X86::XOR16rr, X86::XOR16mr, 0 },
259 { X86::XOR32ri, X86::XOR32mi, 0 },
260 { X86::XOR32ri8, X86::XOR32mi8, 0 },
261 { X86::XOR32rr, X86::XOR32mr, 0 },
262 { X86::XOR64ri32, X86::XOR64mi32, 0 },
263 { X86::XOR64ri8, X86::XOR64mi8, 0 },
264 { X86::XOR64rr, X86::XOR64mr, 0 },
265 { X86::XOR8ri, X86::XOR8mi, 0 },
266 { X86::XOR8rr, X86::XOR8mr, 0 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]267 };
268
269 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]270 unsigned RegOp = OpTbl2Addr[i].RegOp;
271 unsigned MemOp = OpTbl2Addr[i].MemOp;
272 unsigned Flags = OpTbl2Addr[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]273 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
274 RegOp, MemOp,
275 // Index 0, folded load and store, no alignment requirement.
276 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]277 }
278
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]279 static const X86OpTblEntry OpTbl0[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]280 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
281 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
282 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
283 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
284 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]285 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
286 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
287 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
288 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
289 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
290 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
291 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
292 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
293 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
294 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
295 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
296 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
297 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
298 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
299 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
Craig Topperd09a9af2012-12-26 01:47:12 +0000[diff] [blame]300 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]301 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
302 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]303 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
304 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
305 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
306 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
307 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
308 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
309 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
310 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
311 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
312 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
313 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
314 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
315 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
316 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
317 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
318 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
319 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
320 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
321 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
322 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
323 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
324 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]325 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
326 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
327 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
328 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
329 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
330 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]331 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
332 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
333 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
334 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
335 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
336 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
337 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
338 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
339 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
340 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
341 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
342 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
343 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
344 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
345 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
346 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
347 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
348 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
349 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
350 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
351 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
352 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
353 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
354 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
355 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]356 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
357 // AVX 128-bit versions of foldable instructions
Craig Topperd09a9af2012-12-26 01:47:12 +0000[diff] [blame]358 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]359 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
360 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]361 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]362 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
363 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
366 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
367 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
368 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
369 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
370 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
371 // AVX 256-bit foldable instructions
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]372 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]373 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
374 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
377 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]378 };
379
380 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]381 unsigned RegOp = OpTbl0[i].RegOp;
382 unsigned MemOp = OpTbl0[i].MemOp;
383 unsigned Flags = OpTbl0[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]384 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
385 RegOp, MemOp, TB_INDEX_0 | Flags);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]386 }
387
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]388 static const X86OpTblEntry OpTbl1[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]389 { X86::CMP16rr, X86::CMP16rm, 0 },
390 { X86::CMP32rr, X86::CMP32rm, 0 },
391 { X86::CMP64rr, X86::CMP64rm, 0 },
392 { X86::CMP8rr, X86::CMP8rm, 0 },
393 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
394 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
395 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
396 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
397 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
398 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
399 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
400 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
401 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
402 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
403 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
404 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]405 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
406 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
407 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
408 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
409 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
410 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
411 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
412 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]413 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
414 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]415 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
416 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]417 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
418 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
419 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
420 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
421 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
422 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
423 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
424 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]425 { X86::MOV16rr, X86::MOV16rm, 0 },
426 { X86::MOV32rr, X86::MOV32rm, 0 },
427 { X86::MOV64rr, X86::MOV64rm, 0 },
428 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
429 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
430 { X86::MOV8rr, X86::MOV8rm, 0 },
431 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
432 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]433 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
434 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
435 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
436 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]437 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
438 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
439 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
440 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
441 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
442 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
443 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
444 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
445 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
446 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]447 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
448 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
449 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
450 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
451 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
452 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
453 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
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 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]470 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]471 { X86::SQRTSDr, X86::SQRTSDm, 0 },
472 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
473 { X86::SQRTSSr, X86::SQRTSSm, 0 },
474 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
475 { X86::TEST16rr, X86::TEST16rm, 0 },
476 { X86::TEST32rr, X86::TEST32rm, 0 },
477 { X86::TEST64rr, X86::TEST64rm, 0 },
478 { X86::TEST8rr, X86::TEST8rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]479 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]480 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
481 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]482 // AVX 128-bit versions of foldable instructions
483 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
484 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]485 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
486 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]487 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
488 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
Pete Cooper8bbce762012-06-14 22:12:58 +0000[diff] [blame]489 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]490 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
491 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
492 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
493 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
494 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
495 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
496 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
497 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
498 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]499 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
500 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
501 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
502 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
503 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
504 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
505 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
506 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
507 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
508 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
509 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
510 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
Craig Topperb2922162012-12-26 02:14:19 +0000[diff] [blame]511 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]512 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
513 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
514 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
515 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]516 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
517 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
518 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
519 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
520 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
521 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
522 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
523 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
524 { X86::VRCPPSr, X86::VRCPPSm, 0 },
525 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, 0 },
526 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
527 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, 0 },
528 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]529 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]530 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]531 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]532 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
533
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]534 // AVX 256-bit foldable instructions
535 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
536 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]537 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]538 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]539 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]540 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
541 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]542
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]543 // AVX2 foldable instructions
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]544 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
545 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
546 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
547 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
548 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
549 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
550 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
551 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, 0 },
552 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]553 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]554 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]555 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
556 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
Michael Liao2de86af2012-09-26 08:24:51 +0000[diff] [blame]557
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]558 // BMI/BMI2/LZCNT/POPCNT foldable instructions
559 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
560 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
561 { X86::BLSI32rr, X86::BLSI32rm, 0 },
562 { X86::BLSI64rr, X86::BLSI64rm, 0 },
563 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
564 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
565 { X86::BLSR32rr, X86::BLSR32rm, 0 },
566 { X86::BLSR64rr, X86::BLSR64rm, 0 },
567 { X86::BZHI32rr, X86::BZHI32rm, 0 },
568 { X86::BZHI64rr, X86::BZHI64rm, 0 },
569 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
570 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
571 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
572 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
573 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
574 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
Michael Liao2de86af2012-09-26 08:24:51 +0000[diff] [blame]575 { X86::RORX32ri, X86::RORX32mi, 0 },
576 { X86::RORX64ri, X86::RORX64mi, 0 },
Michael Liao2b425e12012-09-26 08:26:25 +0000[diff] [blame]577 { X86::SARX32rr, X86::SARX32rm, 0 },
578 { X86::SARX64rr, X86::SARX64rm, 0 },
579 { X86::SHRX32rr, X86::SHRX32rm, 0 },
580 { X86::SHRX64rr, X86::SHRX64rm, 0 },
581 { X86::SHLX32rr, X86::SHLX32rm, 0 },
582 { X86::SHLX64rr, X86::SHLX64rm, 0 },
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]583 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
584 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
585 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]586 };
587
588 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]589 unsigned RegOp = OpTbl1[i].RegOp;
590 unsigned MemOp = OpTbl1[i].MemOp;
591 unsigned Flags = OpTbl1[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]592 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
593 RegOp, MemOp,
594 // Index 1, folded load
595 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]596 }
597
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]598 static const X86OpTblEntry OpTbl2[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]599 { X86::ADC32rr, X86::ADC32rm, 0 },
600 { X86::ADC64rr, X86::ADC64rm, 0 },
601 { X86::ADD16rr, X86::ADD16rm, 0 },
602 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
603 { X86::ADD32rr, X86::ADD32rm, 0 },
604 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
605 { X86::ADD64rr, X86::ADD64rm, 0 },
606 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
607 { X86::ADD8rr, X86::ADD8rm, 0 },
608 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
609 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
610 { X86::ADDSDrr, X86::ADDSDrm, 0 },
611 { X86::ADDSSrr, X86::ADDSSrm, 0 },
612 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
613 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
614 { X86::AND16rr, X86::AND16rm, 0 },
615 { X86::AND32rr, X86::AND32rm, 0 },
616 { X86::AND64rr, X86::AND64rm, 0 },
617 { X86::AND8rr, X86::AND8rm, 0 },
618 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
619 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
620 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
621 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]622 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
623 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
624 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
625 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]626 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
627 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
628 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
629 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
630 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
631 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
632 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
633 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
634 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
635 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
636 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
637 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
638 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
639 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
640 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
641 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
642 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
643 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
644 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
645 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
646 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
647 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
648 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
649 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
650 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
651 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
652 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
653 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
654 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
655 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
656 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
657 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
658 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
659 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
660 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
661 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
662 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
663 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
664 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
665 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
666 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
667 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
668 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
669 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
670 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
671 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
672 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
673 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
674 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
675 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
676 { X86::CMPSDrr, X86::CMPSDrm, 0 },
677 { X86::CMPSSrr, X86::CMPSSrm, 0 },
678 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
679 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
680 { X86::DIVSDrr, X86::DIVSDrm, 0 },
681 { X86::DIVSSrr, X86::DIVSSrm, 0 },
682 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
683 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
684 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
685 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
686 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
687 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
688 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
689 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
690 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
691 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
692 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
693 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
694 { X86::IMUL16rr, X86::IMUL16rm, 0 },
695 { X86::IMUL32rr, X86::IMUL32rm, 0 },
696 { X86::IMUL64rr, X86::IMUL64rm, 0 },
697 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
698 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
Manman Ren959acb12012-08-13 18:29:41 +0000[diff] [blame]699 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
700 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
701 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
702 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
703 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
704 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]705 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]706 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]707 { X86::MAXSDrr, X86::MAXSDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]708 { X86::MAXSSrr, X86::MAXSSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]709 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]710 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]711 { X86::MINSDrr, X86::MINSDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]712 { X86::MINSSrr, X86::MINSSrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]713 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]714 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
715 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
716 { X86::MULSDrr, X86::MULSDrm, 0 },
717 { X86::MULSSrr, X86::MULSSrm, 0 },
718 { X86::OR16rr, X86::OR16rm, 0 },
719 { X86::OR32rr, X86::OR32rm, 0 },
720 { X86::OR64rr, X86::OR64rm, 0 },
721 { X86::OR8rr, X86::OR8rm, 0 },
722 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
723 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
724 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
725 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]726 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]727 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
728 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
729 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
730 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
731 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
732 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]733 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
734 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]735 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]736 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]737 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
738 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
739 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
740 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]741 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]742 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
743 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]744 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]745 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
746 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
747 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]748 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]749 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]750 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
751 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]752 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]753 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]754 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]755 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]756 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]757 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]758 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
759 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
760 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
761 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
762 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
Benjamin Kramer4669d182012-12-21 14:04:55 +0000[diff] [blame]763 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
764 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
765 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
766 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
767 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
768 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
769 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
770 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]771 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]772 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]773 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
774 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
775 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
776 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
777 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
778 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
779 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]780 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
781 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
782 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
783 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]784 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
785 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
786 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
787 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
788 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
789 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
790 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
791 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
792 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
793 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
794 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
795 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
796 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
797 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
798 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
799 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
800 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
801 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
802 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
803 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
804 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
805 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
806 { X86::SBB32rr, X86::SBB32rm, 0 },
807 { X86::SBB64rr, X86::SBB64rm, 0 },
808 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
809 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
810 { X86::SUB16rr, X86::SUB16rm, 0 },
811 { X86::SUB32rr, X86::SUB32rm, 0 },
812 { X86::SUB64rr, X86::SUB64rm, 0 },
813 { X86::SUB8rr, X86::SUB8rm, 0 },
814 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
815 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
816 { X86::SUBSDrr, X86::SUBSDrm, 0 },
817 { X86::SUBSSrr, X86::SUBSSrm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]818 // FIXME: TEST*rr -> swapped operand of TEST*mr.
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]819 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
820 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
821 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
822 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
823 { X86::XOR16rr, X86::XOR16rm, 0 },
824 { X86::XOR32rr, X86::XOR32rm, 0 },
825 { X86::XOR64rr, X86::XOR64rm, 0 },
826 { X86::XOR8rr, X86::XOR8rm, 0 },
827 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]828 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
829 // AVX 128-bit versions of foldable instructions
830 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
831 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
832 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
833 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
834 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
835 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
836 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
837 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
838 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
839 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
Craig Toppercaef1c52012-12-26 00:35:47 +0000[diff] [blame]840 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
841 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]842 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
843 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]844 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
845 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
846 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]847 { X86::VADDPDrr, X86::VADDPDrm, 0 },
848 { X86::VADDPSrr, X86::VADDPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]849 { X86::VADDSDrr, X86::VADDSDrm, 0 },
850 { X86::VADDSSrr, X86::VADDSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]851 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
852 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
853 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
854 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
855 { X86::VANDPDrr, X86::VANDPDrm, 0 },
856 { X86::VANDPSrr, X86::VANDPSrm, 0 },
857 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
858 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
859 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
860 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
861 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
862 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]863 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
864 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]865 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
866 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]867 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
868 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
869 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
870 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
871 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
872 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
873 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
874 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
875 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
876 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]877 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
878 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
879 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
880 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]881 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
882 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]883 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]884 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]885 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]886 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]887 { X86::VMINPDrr, X86::VMINPDrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]888 { X86::VMINPSrr, X86::VMINPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]889 { X86::VMINSDrr, X86::VMINSDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]890 { X86::VMINSSrr, X86::VMINSSrm, 0 },
Craig Topper81d1e592012-12-26 02:44:47 +0000[diff] [blame]891 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
892 { X86::VMULPDrr, X86::VMULPDrm, 0 },
893 { X86::VMULPSrr, X86::VMULPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]894 { X86::VMULSDrr, X86::VMULSDrm, 0 },
895 { X86::VMULSSrr, X86::VMULSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]896 { X86::VORPDrr, X86::VORPDrm, 0 },
897 { X86::VORPSrr, X86::VORPSrm, 0 },
898 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
899 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
900 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
901 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
902 { X86::VPADDBrr, X86::VPADDBrm, 0 },
903 { X86::VPADDDrr, X86::VPADDDrm, 0 },
904 { X86::VPADDQrr, X86::VPADDQrm, 0 },
905 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
906 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
907 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
908 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
909 { X86::VPADDWrr, X86::VPADDWrm, 0 },
910 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
911 { X86::VPANDNrr, X86::VPANDNrm, 0 },
912 { X86::VPANDrr, X86::VPANDrm, 0 },
913 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
914 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
915 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
916 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
917 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
918 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
919 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
920 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
921 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
922 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
923 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
924 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
925 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
926 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
927 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
928 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
929 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
930 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
931 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
932 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
933 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
934 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
935 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
936 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
937 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
938 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
939 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
940 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
941 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
942 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
943 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
944 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
945 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
946 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
947 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
948 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
949 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
950 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
951 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
952 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
953 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
954 { X86::VPORrr, X86::VPORrm, 0 },
955 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
956 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
957 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
958 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
959 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
960 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
961 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
962 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
963 { X86::VPSRADrr, X86::VPSRADrm, 0 },
964 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
965 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
966 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
967 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
968 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
969 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
970 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
971 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
972 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
973 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
974 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
975 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
976 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
977 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
978 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
979 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
980 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
981 { X86::VPXORrr, X86::VPXORrm, 0 },
982 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
983 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
984 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
985 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]986 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
987 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]988 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
989 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
990 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
991 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
992 { X86::VXORPDrr, X86::VXORPDrm, 0 },
993 { X86::VXORPSrr, X86::VXORPSrm, 0 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]994 // AVX 256-bit foldable instructions
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]995 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
996 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
997 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
998 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
999 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
1000 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
1001 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
1002 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1003 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1004 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1005 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1006 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1007 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1008 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1009 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1010 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1011 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1012 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1013 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1014 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1015 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1016 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1017 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1018 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1019 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1020 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1021 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1022 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1023 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1024 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1025 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1026 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1027 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1028 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1029 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1030 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1031 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1032 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1033 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1034 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1035 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1036 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1037 // AVX2 foldable instructions
Nadav Rotemdc0ad922012-12-24 09:40:33 +0000[diff] [blame]1038 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1039 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1040 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1041 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1042 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1043 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1044 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1045 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1046 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1047 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1048 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1049 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1050 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1051 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1052 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1053 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1054 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1055 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1056 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1057 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1058 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1059 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1060 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1061 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1062 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1063 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1064 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1065 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1066 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1067 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1068 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1069 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
1070 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1071 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
1072 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1073 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1074 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1075 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1076 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1077 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1078 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1079 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1080 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1081 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1082 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1083 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1084 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1085 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1086 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1087 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1088 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1089 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1090 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1091 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1092 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1093 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1094 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1095 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1096 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1097 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1098 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1099 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1100 { X86::VPORYrr, X86::VPORYrm, 0 },
1101 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1102 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1103 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1104 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1105 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1106 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1107 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1108 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1109 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1110 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1111 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1112 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1113 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1114 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1115 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1116 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1117 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1118 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1119 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1120 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1121 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1122 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1123 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1124 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1125 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1126 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1127 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1128 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1129 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1130 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1131 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1132 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1133 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1134 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1135 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1136 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1137 { X86::VPXORYrr, X86::VPXORYrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1138 // FIXME: add AVX 256-bit foldable instructions
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1139
1140 // FMA4 foldable patterns
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1141 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, 0 },
1142 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1143 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1144 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1145 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1146 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1147 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, 0 },
1148 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1149 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1150 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1151 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1152 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1153 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, 0 },
1154 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1155 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1156 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1157 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1158 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1159 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, 0 },
1160 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1161 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1162 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1163 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1164 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1165 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1166 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1167 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1168 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1169 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1170 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1171 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1172 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
Michael Liaof9f7b552012-09-26 08:22:37 +0000[diff] [blame]1173
1174 // BMI/BMI2 foldable instructions
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]1175 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1176 { X86::ANDN64rr, X86::ANDN64rm, 0 },
Michael Liaof9f7b552012-09-26 08:22:37 +0000[diff] [blame]1177 { X86::MULX32rr, X86::MULX32rm, 0 },
1178 { X86::MULX64rr, X86::MULX64rm, 0 },
Craig Topperf924a582012-12-17 05:02:29 +0000[diff] [blame]1179 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1180 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1181 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1182 { X86::PEXT64rr, X86::PEXT64rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1183 };
1184
1185 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]1186 unsigned RegOp = OpTbl2[i].RegOp;
1187 unsigned MemOp = OpTbl2[i].MemOp;
1188 unsigned Flags = OpTbl2[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1189 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1190 RegOp, MemOp,
1191 // Index 2, folded load
1192 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1193 }
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1194
1195 static const X86OpTblEntry OpTbl3[] = {
1196 // FMA foldable instructions
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1197 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1198 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1199 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1200 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1201 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1202 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1203 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, 0 },
1204 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1205
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1206 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1207 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1208 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1209 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1210 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1211 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1212 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1213 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1214 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1215 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1216 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1217 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1218
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1219 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1220 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1221 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1222 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1223 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1224 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1225 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, 0 },
1226 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1227
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1228 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1229 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1230 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1231 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1232 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1233 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1234 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1235 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1236 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1237 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1238 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1239 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1240
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1241 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1242 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1243 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1244 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1245 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1246 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1247 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, 0 },
1248 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1249
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1250 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1251 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1252 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1253 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1254 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1255 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1256 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1257 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1258 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1259 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1260 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1261 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1262
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1263 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1264 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1265 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1266 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1267 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1268 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1269 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, 0 },
1270 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, 0 },
Craig Topper2e127b52012-06-01 05:48:39 +0000[diff] [blame]1271
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1272 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1273 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1274 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1275 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1276 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1277 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1278 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1279 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1280 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1281 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1282 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1283 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1284
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1285 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1286 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1287 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1288 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1289 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1290 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1291 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1292 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1293 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1294 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1295 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1296 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1297
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1298 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1299 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1300 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1301 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1302 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1303 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1304 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1305 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1306 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1307 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1308 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1309 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1310
1311 // FMA4 foldable patterns
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1312 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, 0 },
1313 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1314 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1315 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1316 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1317 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1318 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, 0 },
1319 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1320 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1321 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1322 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1323 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1324 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, 0 },
1325 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1326 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1327 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1328 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1329 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
Craig Topper3b530ea2012-11-04 04:40:08 +0000[diff] [blame]1330 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, 0 },
1331 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, 0 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1332 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1333 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1334 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1335 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1336 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1337 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1338 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1339 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1340 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1341 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1342 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1343 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1344 };
1345
1346 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1347 unsigned RegOp = OpTbl3[i].RegOp;
1348 unsigned MemOp = OpTbl3[i].MemOp;
1349 unsigned Flags = OpTbl3[i].Flags;
1350 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1351 RegOp, MemOp,
1352 // Index 3, folded load
1353 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1354 }
1355
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]1356}
1357
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1358void
1359X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1360 MemOp2RegOpTableType &M2RTable,
1361 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1362 if ((Flags & TB_NO_FORWARD) == 0) {
1363 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1364 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1365 }
1366 if ((Flags & TB_NO_REVERSE) == 0) {
1367 assert(!M2RTable.count(MemOp) &&
1368 "Duplicated entries in unfolding maps?");
1369 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1370 }
1371}
1372
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1373bool
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1374X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1375 unsigned &SrcReg, unsigned &DstReg,
1376 unsigned &SubIdx) const {
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1377 switch (MI.getOpcode()) {
1378 default: break;
1379 case X86::MOVSX16rr8:
1380 case X86::MOVZX16rr8:
1381 case X86::MOVSX32rr8:
1382 case X86::MOVZX32rr8:
1383 case X86::MOVSX64rr8:
1384 case X86::MOVZX64rr8:
Evan Chengceb5a4e2010-01-13 08:01:32 +0000[diff] [blame]1385 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1386 // It's not always legal to reference the low 8-bit of the larger
1387 // register in 32-bit mode.
1388 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1389 case X86::MOVSX32rr16:
1390 case X86::MOVZX32rr16:
1391 case X86::MOVSX64rr16:
1392 case X86::MOVZX64rr16:
1393 case X86::MOVSX64rr32:
1394 case X86::MOVZX64rr32: {
1395 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1396 // Be conservative.
1397 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1398 SrcReg = MI.getOperand(1).getReg();
1399 DstReg = MI.getOperand(0).getReg();
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1400 switch (MI.getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1401 default: llvm_unreachable("Unreachable!");
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1402 case X86::MOVSX16rr8:
1403 case X86::MOVZX16rr8:
1404 case X86::MOVSX32rr8:
1405 case X86::MOVZX32rr8:
1406 case X86::MOVSX64rr8:
1407 case X86::MOVZX64rr8:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1408 SubIdx = X86::sub_8bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1409 break;
1410 case X86::MOVSX32rr16:
1411 case X86::MOVZX32rr16:
1412 case X86::MOVSX64rr16:
1413 case X86::MOVZX64rr16:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1414 SubIdx = X86::sub_16bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1415 break;
1416 case X86::MOVSX64rr32:
1417 case X86::MOVZX64rr32:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1418 SubIdx = X86::sub_32bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1419 break;
1420 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1421 return true;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1422 }
1423 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1424 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1425}
1426
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1427/// isFrameOperand - Return true and the FrameIndex if the specified
1428/// operand and follow operands form a reference to the stack frame.
1429bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1430 int &FrameIndex) const {
1431 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1432 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1433 MI->getOperand(Op+1).getImm() == 1 &&
1434 MI->getOperand(Op+2).getReg() == 0 &&
1435 MI->getOperand(Op+3).getImm() == 0) {
1436 FrameIndex = MI->getOperand(Op).getIndex();
1437 return true;
1438 }
1439 return false;
1440}
1441
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1442static bool isFrameLoadOpcode(int Opcode) {
1443 switch (Opcode) {
David Blaikie46a9f012012-01-20 21:51:11 +0000[diff] [blame]1444 default:
1445 return false;
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1446 case X86::MOV8rm:
1447 case X86::MOV16rm:
1448 case X86::MOV32rm:
Evan Cheng11b0a5d2006-09-08 06:48:29 +0000[diff] [blame]1449 case X86::MOV64rm:
Dale Johannesen3d7008c2007-07-04 21:07:47 +0000[diff] [blame]1450 case X86::LD_Fp64m:
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1451 case X86::MOVSSrm:
1452 case X86::MOVSDrm:
Chris Lattnerbfc2c682006-04-18 16:44:51 +0000[diff] [blame]1453 case X86::MOVAPSrm:
1454 case X86::MOVAPDrm:
Dan Gohmanbdc0f8b2009-01-09 02:40:34 +0000[diff] [blame]1455 case X86::MOVDQArm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1456 case X86::VMOVSSrm:
1457 case X86::VMOVSDrm:
1458 case X86::VMOVAPSrm:
1459 case X86::VMOVAPDrm:
1460 case X86::VMOVDQArm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1461 case X86::VMOVAPSYrm:
1462 case X86::VMOVAPDYrm:
1463 case X86::VMOVDQAYrm:
Bill Wendlinge7b2a862007-04-03 06:00:37 +0000[diff] [blame]1464 case X86::MMX_MOVD64rm:
1465 case X86::MMX_MOVQ64rm:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1466 return true;
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1467 }
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1468}
1469
1470static bool isFrameStoreOpcode(int Opcode) {
1471 switch (Opcode) {
1472 default: break;
1473 case X86::MOV8mr:
1474 case X86::MOV16mr:
1475 case X86::MOV32mr:
1476 case X86::MOV64mr:
1477 case X86::ST_FpP64m:
1478 case X86::MOVSSmr:
1479 case X86::MOVSDmr:
1480 case X86::MOVAPSmr:
1481 case X86::MOVAPDmr:
1482 case X86::MOVDQAmr:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1483 case X86::VMOVSSmr:
1484 case X86::VMOVSDmr:
1485 case X86::VMOVAPSmr:
1486 case X86::VMOVAPDmr:
1487 case X86::VMOVDQAmr:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1488 case X86::VMOVAPSYmr:
1489 case X86::VMOVAPDYmr:
1490 case X86::VMOVDQAYmr:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1491 case X86::MMX_MOVD64mr:
1492 case X86::MMX_MOVQ64mr:
1493 case X86::MMX_MOVNTQmr:
1494 return true;
1495 }
1496 return false;
1497}
1498
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1499unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1500 int &FrameIndex) const {
1501 if (isFrameLoadOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1502 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1503 return MI->getOperand(0).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1504 return 0;
1505}
1506
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1507unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1508 int &FrameIndex) const {
1509 if (isFrameLoadOpcode(MI->getOpcode())) {
1510 unsigned Reg;
1511 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1512 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1513 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1514 const MachineMemOperand *Dummy;
1515 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1516 }
1517 return 0;
1518}
1519
Dan Gohman0b273252008-11-18 19:49:32 +0000[diff] [blame]1520unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1521 int &FrameIndex) const {
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1522 if (isFrameStoreOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1523 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1524 isFrameOperand(MI, 0, FrameIndex))
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]1525 return MI->getOperand(X86::AddrNumOperands).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1526 return 0;
1527}
1528
1529unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1530 int &FrameIndex) const {
1531 if (isFrameStoreOpcode(MI->getOpcode())) {
1532 unsigned Reg;
1533 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1534 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1535 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1536 const MachineMemOperand *Dummy;
1537 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1538 }
1539 return 0;
1540}
1541
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1542/// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1543/// X86::MOVPC32r.
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1544static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
Jakob Stoklund Olesen3b9a4422012-08-08 00:40:47 +0000[diff] [blame]1545 // Don't waste compile time scanning use-def chains of physregs.
1546 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1547 return false;
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1548 bool isPICBase = false;
1549 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1550 E = MRI.def_end(); I != E; ++I) {
1551 MachineInstr *DefMI = I.getOperand().getParent();
1552 if (DefMI->getOpcode() != X86::MOVPC32r)
1553 return false;
1554 assert(!isPICBase && "More than one PIC base?");
1555 isPICBase = true;
1556 }
1557 return isPICBase;
1558}
Evan Cheng1973a462008-03-31 07:54:19 +0000[diff] [blame]1559
Bill Wendling1e117682008-05-12 20:54:26 +0000[diff] [blame]1560bool
Dan Gohmane919de52009-10-10 00:34:18 +0000[diff] [blame]1561X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1562 AliasAnalysis *AA) const {
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1563 switch (MI->getOpcode()) {
1564 default: break;
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1565 case X86::MOV8rm:
1566 case X86::MOV16rm:
1567 case X86::MOV32rm:
1568 case X86::MOV64rm:
1569 case X86::LD_Fp64m:
1570 case X86::MOVSSrm:
1571 case X86::MOVSDrm:
1572 case X86::MOVAPSrm:
1573 case X86::MOVUPSrm:
1574 case X86::MOVAPDrm:
1575 case X86::MOVDQArm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1576 case X86::MOVDQUrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1577 case X86::VMOVSSrm:
1578 case X86::VMOVSDrm:
1579 case X86::VMOVAPSrm:
1580 case X86::VMOVUPSrm:
1581 case X86::VMOVAPDrm:
1582 case X86::VMOVDQArm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1583 case X86::VMOVDQUrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1584 case X86::VMOVAPSYrm:
1585 case X86::VMOVUPSYrm:
1586 case X86::VMOVAPDYrm:
1587 case X86::VMOVDQAYrm:
Craig Topper922f10a2012-12-06 06:49:16 +0000[diff] [blame]1588 case X86::VMOVDQUYrm:
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1589 case X86::MMX_MOVD64rm:
1590 case X86::MMX_MOVQ64rm:
1591 case X86::FsVMOVAPSrm:
1592 case X86::FsVMOVAPDrm:
1593 case X86::FsMOVAPSrm:
1594 case X86::FsMOVAPDrm: {
1595 // Loads from constant pools are trivially rematerializable.
1596 if (MI->getOperand(1).isReg() &&
1597 MI->getOperand(2).isImm() &&
1598 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1599 MI->isInvariantLoad(AA)) {
1600 unsigned BaseReg = MI->getOperand(1).getReg();
1601 if (BaseReg == 0 || BaseReg == X86::RIP)
1602 return true;
1603 // Allow re-materialization of PIC load.
1604 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1605 return false;
1606 const MachineFunction &MF = *MI->getParent()->getParent();
1607 const MachineRegisterInfo &MRI = MF.getRegInfo();
1608 return regIsPICBase(BaseReg, MRI);
Evan Cheng94ba37f2008-02-22 09:25:47 +0000[diff] [blame]1609 }
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1610 return false;
1611 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1612
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1613 case X86::LEA32r:
1614 case X86::LEA64r: {
1615 if (MI->getOperand(2).isImm() &&
1616 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1617 !MI->getOperand(4).isReg()) {
1618 // lea fi#, lea GV, etc. are all rematerializable.
1619 if (!MI->getOperand(1).isReg())
1620 return true;
1621 unsigned BaseReg = MI->getOperand(1).getReg();
1622 if (BaseReg == 0)
1623 return true;
1624 // Allow re-materialization of lea PICBase + x.
1625 const MachineFunction &MF = *MI->getParent()->getParent();
1626 const MachineRegisterInfo &MRI = MF.getRegInfo();
1627 return regIsPICBase(BaseReg, MRI);
1628 }
1629 return false;
1630 }
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1631 }
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1632
Dan Gohmane8c1e422007-06-26 00:48:07 +0000[diff] [blame]1633 // All other instructions marked M_REMATERIALIZABLE are always trivially
1634 // rematerializable.
1635 return true;
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1636}
1637
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1638/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1639/// would clobber the EFLAGS condition register. Note the result may be
1640/// conservative. If it cannot definitely determine the safety after visiting
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1641/// a few instructions in each direction it assumes it's not safe.
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1642static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1643 MachineBasicBlock::iterator I) {
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1644 MachineBasicBlock::iterator E = MBB.end();
1645
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1646 // For compile time consideration, if we are not able to determine the
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1647 // safety after visiting 4 instructions in each direction, we will assume
1648 // it's not safe.
1649 MachineBasicBlock::iterator Iter = I;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1650 for (unsigned i = 0; Iter != E && i < 4; ++i) {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1651 bool SeenDef = false;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1652 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1653 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1654 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1655 SeenDef = true;
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1656 if (!MO.isReg())
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1657 continue;
1658 if (MO.getReg() == X86::EFLAGS) {
1659 if (MO.isUse())
1660 return false;
1661 SeenDef = true;
1662 }
1663 }
1664
1665 if (SeenDef)
1666 // This instruction defines EFLAGS, no need to look any further.
1667 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1668 ++Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1669 // Skip over DBG_VALUE.
1670 while (Iter != E && Iter->isDebugValue())
1671 ++Iter;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1672 }
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1673
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1674 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1675 // live in.
1676 if (Iter == E) {
1677 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1678 SE = MBB.succ_end(); SI != SE; ++SI)
1679 if ((*SI)->isLiveIn(X86::EFLAGS))
1680 return false;
1681 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1682 }
1683
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1684 MachineBasicBlock::iterator B = MBB.begin();
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1685 Iter = I;
1686 for (unsigned i = 0; i < 4; ++i) {
1687 // If we make it to the beginning of the block, it's safe to clobber
Sylvestre Ledru91ce36c2012-09-27 10:14:43 +0000[diff] [blame]1688 // EFLAGS iff EFLAGS is not live-in.
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1689 if (Iter == B)
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1690 return !MBB.isLiveIn(X86::EFLAGS);
1691
1692 --Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1693 // Skip over DBG_VALUE.
1694 while (Iter != B && Iter->isDebugValue())
1695 --Iter;
1696
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1697 bool SawKill = false;
1698 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1699 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1700 // A register mask may clobber EFLAGS, but we should still look for a
1701 // live EFLAGS def.
1702 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1703 SawKill = true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1704 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1705 if (MO.isDef()) return MO.isDead();
1706 if (MO.isKill()) SawKill = true;
1707 }
1708 }
1709
1710 if (SawKill)
1711 // This instruction kills EFLAGS and doesn't redefine it, so
1712 // there's no need to look further.
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1713 return true;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1714 }
1715
1716 // Conservative answer.
1717 return false;
1718}
1719
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1720void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1721 MachineBasicBlock::iterator I,
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1722 unsigned DestReg, unsigned SubIdx,
Evan Cheng6ad7da92009-11-14 02:55:43 +0000[diff] [blame]1723 const MachineInstr *Orig,
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1724 const TargetRegisterInfo &TRI) const {
Dan Gohman90c600d2010-05-07 01:28:10 +0000[diff] [blame]1725 DebugLoc DL = Orig->getDebugLoc();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1726
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1727 // MOV32r0 etc. are implemented with xor which clobbers condition code.
1728 // Re-materialize them as movri instructions to avoid side effects.
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1729 bool Clone = true;
1730 unsigned Opc = Orig->getOpcode();
1731 switch (Opc) {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1732 default: break;
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1733 case X86::MOV8r0:
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]1734 case X86::MOV16r0:
1735 case X86::MOV32r0:
1736 case X86::MOV64r0: {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1737 if (!isSafeToClobberEFLAGS(MBB, I)) {
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1738 switch (Opc) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1739 default: llvm_unreachable("Unreachable!");
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1740 case X86::MOV8r0: Opc = X86::MOV8ri; break;
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]1741 case X86::MOV16r0: Opc = X86::MOV16ri; break;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1742 case X86::MOV32r0: Opc = X86::MOV32ri; break;
Dan Gohman952f6f92010-02-26 16:49:27 +0000[diff] [blame]1743 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1744 }
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1745 Clone = false;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1746 }
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1747 break;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1748 }
1749 }
1750
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1751 if (Clone) {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1752 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1753 MBB.insert(I, MI);
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1754 } else {
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1755 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1756 }
Evan Cheng147cb762008-04-16 23:44:44 +0000[diff] [blame]1757
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1758 MachineInstr *NewMI = prior(I);
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1759 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1760}
1761
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1762/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1763/// is not marked dead.
1764static bool hasLiveCondCodeDef(MachineInstr *MI) {
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1765 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1766 MachineOperand &MO = MI->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1767 if (MO.isReg() && MO.isDef() &&
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1768 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1769 return true;
1770 }
1771 }
1772 return false;
1773}
1774
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame^]1775/// getTruncatedShiftCount - check whether the shift count for a machine operand
1776/// is non-zero.
1777inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
1778 unsigned ShiftAmtOperandIdx) {
1779 // The shift count is six bits with the REX.W prefix and five bits without.
1780 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
1781 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
1782 return Imm & ShiftCountMask;
1783}
1784
1785/// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
1786/// can be represented by a LEA instruction.
1787inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
1788 // Left shift instructions can be transformed into load-effective-address
1789 // instructions if we can encode them appropriately.
1790 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
1791 // The SIB.scale field is two bits wide which means that we can encode any
1792 // shift amount less than 4.
1793 return ShAmt < 4 && ShAmt > 0;
1794}
1795
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1796/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1797/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1798/// to a 32-bit superregister and then truncating back down to a 16-bit
1799/// subregister.
1800MachineInstr *
1801X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1802 MachineFunction::iterator &MFI,
1803 MachineBasicBlock::iterator &MBBI,
1804 LiveVariables *LV) const {
1805 MachineInstr *MI = MBBI;
1806 unsigned Dest = MI->getOperand(0).getReg();
1807 unsigned Src = MI->getOperand(1).getReg();
1808 bool isDead = MI->getOperand(0).isDead();
1809 bool isKill = MI->getOperand(1).isKill();
1810
1811 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1812 ? X86::LEA64_32r : X86::LEA32r;
1813 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1814 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1815 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1816
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1817 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1818 // well be shifting and then extracting the lower 16-bits.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1819 // This has the potential to cause partial register stall. e.g.
Evan Cheng3974c8d2009-12-12 18:55:26 +0000[diff] [blame]1820 // movw (%rbp,%rcx,2), %dx
1821 // leal -65(%rdx), %esi
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1822 // But testing has shown this *does* help performance in 64-bit mode (at
1823 // least on modern x86 machines).
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1824 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1825 MachineInstr *InsMI =
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1826 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1827 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1828 .addReg(Src, getKillRegState(isKill));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1829
1830 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1831 get(Opc), leaOutReg);
1832 switch (MIOpc) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1833 default: llvm_unreachable("Unreachable!");
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1834 case X86::SHL16ri: {
1835 unsigned ShAmt = MI->getOperand(2).getImm();
1836 MIB.addReg(0).addImm(1 << ShAmt)
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1837 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1838 break;
1839 }
1840 case X86::INC16r:
1841 case X86::INC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1842 addRegOffset(MIB, leaInReg, true, 1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1843 break;
1844 case X86::DEC16r:
1845 case X86::DEC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1846 addRegOffset(MIB, leaInReg, true, -1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1847 break;
1848 case X86::ADD16ri:
1849 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]1850 case X86::ADD16ri_DB:
1851 case X86::ADD16ri8_DB:
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1852 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1853 break;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1854 case X86::ADD16rr:
1855 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1856 unsigned Src2 = MI->getOperand(2).getReg();
1857 bool isKill2 = MI->getOperand(2).isKill();
1858 unsigned leaInReg2 = 0;
1859 MachineInstr *InsMI2 = 0;
1860 if (Src == Src2) {
1861 // ADD16rr %reg1028<kill>, %reg1028
1862 // just a single insert_subreg.
1863 addRegReg(MIB, leaInReg, true, leaInReg, false);
1864 } else {
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1865 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1866 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1867 // well be shifting and then extracting the lower 16-bits.
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1868 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1869 InsMI2 =
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1870 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1871 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1872 .addReg(Src2, getKillRegState(isKill2));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1873 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1874 }
1875 if (LV && isKill2 && InsMI2)
1876 LV->replaceKillInstruction(Src2, MI, InsMI2);
1877 break;
1878 }
1879 }
1880
1881 MachineInstr *NewMI = MIB;
1882 MachineInstr *ExtMI =
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1883 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1884 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1885 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1886
1887 if (LV) {
1888 // Update live variables
1889 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1890 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1891 if (isKill)
1892 LV->replaceKillInstruction(Src, MI, InsMI);
1893 if (isDead)
1894 LV->replaceKillInstruction(Dest, MI, ExtMI);
1895 }
1896
1897 return ExtMI;
1898}
1899
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1900/// convertToThreeAddress - This method must be implemented by targets that
1901/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1902/// may be able to convert a two-address instruction into a true
1903/// three-address instruction on demand. This allows the X86 target (for
1904/// example) to convert ADD and SHL instructions into LEA instructions if they
1905/// would require register copies due to two-addressness.
1906///
1907/// This method returns a null pointer if the transformation cannot be
1908/// performed, otherwise it returns the new instruction.
1909///
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1910MachineInstr *
1911X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1912 MachineBasicBlock::iterator &MBBI,
Owen Anderson30cc0282008-07-02 23:41:07 +0000[diff] [blame]1913 LiveVariables *LV) const {
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1914 MachineInstr *MI = MBBI;
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame^]1915
1916 // The following opcodes also sets the condition code register(s). Only
1917 // convert them to equivalent lea if the condition code register def's
1918 // are dead!
1919 if (hasLiveCondCodeDef(MI))
1920 return 0;
1921
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1922 MachineFunction &MF = *MI->getParent()->getParent();
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1923 // All instructions input are two-addr instructions. Get the known operands.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1924 const MachineOperand &Dest = MI->getOperand(0);
1925 const MachineOperand &Src = MI->getOperand(1);
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1926
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]1927 MachineInstr *NewMI = NULL;
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1928 // 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]1929 // we have better subtarget support, enable the 16-bit LEA generation here.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1930 // 16-bit LEA is also slow on Core2.
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1931 bool DisableLEA16 = true;
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1932 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1933
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1934 unsigned MIOpc = MI->getOpcode();
1935 switch (MIOpc) {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1936 case X86::SHUFPSrri: {
1937 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1938 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1939
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]1940 unsigned B = MI->getOperand(1).getReg();
1941 unsigned C = MI->getOperand(2).getReg();
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1942 if (B != C) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1943 unsigned M = MI->getOperand(3).getImm();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1944 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1945 .addOperand(Dest).addOperand(Src).addImm(M);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1946 break;
1947 }
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1948 case X86::SHUFPDrri: {
1949 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1950 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1951
1952 unsigned B = MI->getOperand(1).getReg();
1953 unsigned C = MI->getOperand(2).getReg();
1954 if (B != C) return 0;
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1955 unsigned M = MI->getOperand(3).getImm();
1956
1957 // Convert to PSHUFD mask.
1958 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1959
1960 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1961 .addOperand(Dest).addOperand(Src).addImm(M);
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1962 break;
1963 }
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1964 case X86::SHL64ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1965 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame^]1966 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1967 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1968
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1969 // LEA can't handle RSP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1970 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1971 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1972 &X86::GR64_NOSPRegClass))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1973 return 0;
1974
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1975 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1976 .addOperand(Dest)
1977 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1978 break;
1979 }
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1980 case X86::SHL32ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1981 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame^]1982 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1983 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1984
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1985 // LEA can't handle ESP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1986 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1987 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1988 &X86::GR32_NOSPRegClass))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1989 return 0;
1990
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1991 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1992 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1993 .addOperand(Dest)
1994 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1995 break;
1996 }
1997 case X86::SHL16ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1998 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame^]1999 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2000 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2001
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2002 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2003 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2004 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2005 .addOperand(Dest)
2006 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]2007 break;
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]2008 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2009 default: {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]2010
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2011 switch (MIOpc) {
2012 default: return 0;
2013 case X86::INC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]2014 case X86::INC32r:
2015 case X86::INC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2016 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]2017 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2018 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]2019 const TargetRegisterClass *RC = MIOpc == X86::INC64r ?
2020 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
2021 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2022
2023 // LEA can't handle RSP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2024 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2025 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2026 return 0;
2027
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2028 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2029 .addOperand(Dest).addOperand(Src), 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2030 break;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2031 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2032 case X86::INC16r:
2033 case X86::INC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2034 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2035 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2036 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2037 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2038 .addOperand(Dest).addOperand(Src), 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2039 break;
2040 case X86::DEC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]2041 case X86::DEC32r:
2042 case X86::DEC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2043 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]2044 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2045 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]2046 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ?
2047 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
2048 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2049 // LEA can't handle RSP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2050 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2051 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2052 return 0;
2053
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2054 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2055 .addOperand(Dest).addOperand(Src), -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2056 break;
2057 }
2058 case X86::DEC16r:
2059 case X86::DEC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2060 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2061 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2062 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2063 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2064 .addOperand(Dest).addOperand(Src), -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2065 break;
2066 case X86::ADD64rr:
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2067 case X86::ADD64rr_DB:
2068 case X86::ADD32rr:
2069 case X86::ADD32rr_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2070 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2071 unsigned Opc;
Craig Topper760b1342012-02-22 05:59:10 +0000[diff] [blame]2072 const TargetRegisterClass *RC;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2073 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
2074 Opc = X86::LEA64r;
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]2075 RC = &X86::GR64_NOSPRegClass;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2076 } else {
2077 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]2078 RC = &X86::GR32_NOSPRegClass;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2079 }
2080
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2081
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2082 unsigned Src2 = MI->getOperand(2).getReg();
2083 bool isKill2 = MI->getOperand(2).isKill();
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2084
2085 // LEA can't handle RSP.
2086 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2087 !MF.getRegInfo().constrainRegClass(Src2, RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2088 return 0;
2089
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2090 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2091 .addOperand(Dest),
2092 Src.getReg(), Src.isKill(), Src2, isKill2);
Nadav Rotem4968e452012-07-16 10:52:25 +0000[diff] [blame]2093
2094 // Preserve undefness of the operands.
2095 bool isUndef = MI->getOperand(1).isUndef();
2096 bool isUndef2 = MI->getOperand(2).isUndef();
2097 NewMI->getOperand(1).setIsUndef(isUndef);
2098 NewMI->getOperand(3).setIsUndef(isUndef2);
2099
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2100 if (LV && isKill2)
2101 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2102 break;
2103 }
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2104 case X86::ADD16rr:
2105 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2106 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2107 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2108 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2109 unsigned Src2 = MI->getOperand(2).getReg();
2110 bool isKill2 = MI->getOperand(2).isKill();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2111 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2112 .addOperand(Dest),
2113 Src.getReg(), Src.isKill(), Src2, isKill2);
2114
2115 // Preserve undefness of the operands.
2116 bool isUndef = MI->getOperand(1).isUndef();
2117 bool isUndef2 = MI->getOperand(2).isUndef();
2118 NewMI->getOperand(1).setIsUndef(isUndef);
2119 NewMI->getOperand(3).setIsUndef(isUndef2);
2120
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2121 if (LV && isKill2)
2122 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2123 break;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2124 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2125 case X86::ADD64ri32:
2126 case X86::ADD64ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2127 case X86::ADD64ri32_DB:
2128 case X86::ADD64ri8_DB:
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2129 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2130 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2131 .addOperand(Dest).addOperand(Src),
2132 MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2133 break;
2134 case X86::ADD32ri:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2135 case X86::ADD32ri8:
2136 case X86::ADD32ri_DB:
2137 case X86::ADD32ri8_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2138 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2139 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2140 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2141 .addOperand(Dest).addOperand(Src),
2142 MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2143 break;
2144 }
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2145 case X86::ADD16ri:
2146 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2147 case X86::ADD16ri_DB:
2148 case X86::ADD16ri8_DB:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2149 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2150 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2151 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2152 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2153 .addOperand(Dest).addOperand(Src),
2154 MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2155 break;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2156 }
2157 }
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2158 }
2159
Evan Cheng1bc1cae2008-02-07 08:29:53 +0000[diff] [blame]2160 if (!NewMI) return 0;
2161
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2162 if (LV) { // Update live variables
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2163 if (Src.isKill())
2164 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2165 if (Dest.isDead())
2166 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2167 }
2168
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2169 MFI->insert(MBBI, NewMI); // Insert the new inst
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]2170 return NewMI;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2171}
2172
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2173/// commuteInstruction - We have a few instructions that must be hacked on to
2174/// commute them.
2175///
Evan Cheng03553bb2008-06-16 07:33:11 +0000[diff] [blame]2176MachineInstr *
2177X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2178 switch (MI->getOpcode()) {
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2179 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2180 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2181 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2182 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2183 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2184 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2185 unsigned Opc;
2186 unsigned Size;
2187 switch (MI->getOpcode()) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2188 default: llvm_unreachable("Unreachable!");
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2189 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2190 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2191 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2192 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2193 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2194 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2195 }
Chris Lattner5c463782007-12-30 20:49:49 +0000[diff] [blame]2196 unsigned Amt = MI->getOperand(3).getImm();
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2197 if (NewMI) {
2198 MachineFunction &MF = *MI->getParent()->getParent();
2199 MI = MF.CloneMachineInstr(MI);
2200 NewMI = false;
Evan Cheng244183e2008-02-13 02:46:49 +0000[diff] [blame]2201 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2202 MI->setDesc(get(Opc));
2203 MI->getOperand(3).setImm(Size-Amt);
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]2204 return TargetInstrInfo::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2205 }
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2206 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2207 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2208 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2209 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2210 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2211 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2212 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2213 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2214 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2215 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2216 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2217 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2218 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2219 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2220 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2221 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2222 unsigned Opc;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2223 switch (MI->getOpcode()) {
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2224 default: llvm_unreachable("Unreachable!");
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2225 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2226 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2227 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2228 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2229 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2230 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2231 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2232 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2233 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2234 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2235 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2236 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
Chris Lattner1a1c6002010-10-05 23:00:14 +0000[diff] [blame]2237 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2238 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2239 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2240 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2241 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2242 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2243 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2244 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2245 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2246 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2247 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2248 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2249 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2250 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2251 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2252 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2253 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2254 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2255 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2256 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2257 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2258 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2259 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2260 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2261 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2262 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2263 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2264 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2265 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2266 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2267 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2268 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2269 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2270 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2271 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2272 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2273 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2274 if (NewMI) {
2275 MachineFunction &MF = *MI->getParent()->getParent();
2276 MI = MF.CloneMachineInstr(MI);
2277 NewMI = false;
2278 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]2279 MI->setDesc(get(Opc));
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2280 // Fallthrough intended.
2281 }
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2282 default:
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]2283 return TargetInstrInfo::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2284 }
2285}
2286
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2287static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2288 switch (BrOpc) {
2289 default: return X86::COND_INVALID;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2290 case X86::JE_4: return X86::COND_E;
2291 case X86::JNE_4: return X86::COND_NE;
2292 case X86::JL_4: return X86::COND_L;
2293 case X86::JLE_4: return X86::COND_LE;
2294 case X86::JG_4: return X86::COND_G;
2295 case X86::JGE_4: return X86::COND_GE;
2296 case X86::JB_4: return X86::COND_B;
2297 case X86::JBE_4: return X86::COND_BE;
2298 case X86::JA_4: return X86::COND_A;
2299 case X86::JAE_4: return X86::COND_AE;
2300 case X86::JS_4: return X86::COND_S;
2301 case X86::JNS_4: return X86::COND_NS;
2302 case X86::JP_4: return X86::COND_P;
2303 case X86::JNP_4: return X86::COND_NP;
2304 case X86::JO_4: return X86::COND_O;
2305 case X86::JNO_4: return X86::COND_NO;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2306 }
2307}
2308
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2309/// getCondFromSETOpc - return condition code of a SET opcode.
2310static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2311 switch (Opc) {
2312 default: return X86::COND_INVALID;
2313 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2314 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2315 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2316 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2317 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2318 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2319 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2320 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2321 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2322 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2323 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2324 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2325 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2326 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2327 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2328 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2329 }
2330}
2331
2332/// getCondFromCmovOpc - return condition code of a CMov opcode.
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]2333X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2334 switch (Opc) {
2335 default: return X86::COND_INVALID;
2336 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2337 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2338 return X86::COND_A;
2339 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2340 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2341 return X86::COND_AE;
2342 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2343 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2344 return X86::COND_B;
2345 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2346 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2347 return X86::COND_BE;
2348 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2349 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2350 return X86::COND_E;
2351 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2352 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2353 return X86::COND_G;
2354 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2355 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2356 return X86::COND_GE;
2357 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2358 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2359 return X86::COND_L;
2360 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2361 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2362 return X86::COND_LE;
2363 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2364 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2365 return X86::COND_NE;
2366 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2367 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2368 return X86::COND_NO;
2369 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2370 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2371 return X86::COND_NP;
2372 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2373 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2374 return X86::COND_NS;
2375 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2376 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2377 return X86::COND_O;
2378 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2379 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2380 return X86::COND_P;
2381 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2382 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2383 return X86::COND_S;
2384 }
2385}
2386
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2387unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2388 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2389 default: llvm_unreachable("Illegal condition code!");
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2390 case X86::COND_E: return X86::JE_4;
2391 case X86::COND_NE: return X86::JNE_4;
2392 case X86::COND_L: return X86::JL_4;
2393 case X86::COND_LE: return X86::JLE_4;
2394 case X86::COND_G: return X86::JG_4;
2395 case X86::COND_GE: return X86::JGE_4;
2396 case X86::COND_B: return X86::JB_4;
2397 case X86::COND_BE: return X86::JBE_4;
2398 case X86::COND_A: return X86::JA_4;
2399 case X86::COND_AE: return X86::JAE_4;
2400 case X86::COND_S: return X86::JS_4;
2401 case X86::COND_NS: return X86::JNS_4;
2402 case X86::COND_P: return X86::JP_4;
2403 case X86::COND_NP: return X86::JNP_4;
2404 case X86::COND_O: return X86::JO_4;
2405 case X86::COND_NO: return X86::JNO_4;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2406 }
2407}
2408
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2409/// GetOppositeBranchCondition - Return the inverse of the specified condition,
2410/// e.g. turning COND_E to COND_NE.
2411X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2412 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2413 default: llvm_unreachable("Illegal condition code!");
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2414 case X86::COND_E: return X86::COND_NE;
2415 case X86::COND_NE: return X86::COND_E;
2416 case X86::COND_L: return X86::COND_GE;
2417 case X86::COND_LE: return X86::COND_G;
2418 case X86::COND_G: return X86::COND_LE;
2419 case X86::COND_GE: return X86::COND_L;
2420 case X86::COND_B: return X86::COND_AE;
2421 case X86::COND_BE: return X86::COND_A;
2422 case X86::COND_A: return X86::COND_BE;
2423 case X86::COND_AE: return X86::COND_B;
2424 case X86::COND_S: return X86::COND_NS;
2425 case X86::COND_NS: return X86::COND_S;
2426 case X86::COND_P: return X86::COND_NP;
2427 case X86::COND_NP: return X86::COND_P;
2428 case X86::COND_O: return X86::COND_NO;
2429 case X86::COND_NO: return X86::COND_O;
2430 }
2431}
2432
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2433/// getSwappedCondition - assume the flags are set by MI(a,b), return
2434/// the condition code if we modify the instructions such that flags are
2435/// set by MI(b,a).
Benjamin Kramerabbfe692012-07-13 13:25:15 +0000[diff] [blame]2436static X86::CondCode getSwappedCondition(X86::CondCode CC) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2437 switch (CC) {
2438 default: return X86::COND_INVALID;
2439 case X86::COND_E: return X86::COND_E;
2440 case X86::COND_NE: return X86::COND_NE;
2441 case X86::COND_L: return X86::COND_G;
2442 case X86::COND_LE: return X86::COND_GE;
2443 case X86::COND_G: return X86::COND_L;
2444 case X86::COND_GE: return X86::COND_LE;
2445 case X86::COND_B: return X86::COND_A;
2446 case X86::COND_BE: return X86::COND_AE;
2447 case X86::COND_A: return X86::COND_B;
2448 case X86::COND_AE: return X86::COND_BE;
2449 }
2450}
2451
2452/// getSETFromCond - Return a set opcode for the given condition and
2453/// whether it has memory operand.
2454static unsigned getSETFromCond(X86::CondCode CC,
2455 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2456 static const uint16_t Opc[16][2] = {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2457 { X86::SETAr, X86::SETAm },
2458 { X86::SETAEr, X86::SETAEm },
2459 { X86::SETBr, X86::SETBm },
2460 { X86::SETBEr, X86::SETBEm },
2461 { X86::SETEr, X86::SETEm },
2462 { X86::SETGr, X86::SETGm },
2463 { X86::SETGEr, X86::SETGEm },
2464 { X86::SETLr, X86::SETLm },
2465 { X86::SETLEr, X86::SETLEm },
2466 { X86::SETNEr, X86::SETNEm },
2467 { X86::SETNOr, X86::SETNOm },
2468 { X86::SETNPr, X86::SETNPm },
2469 { X86::SETNSr, X86::SETNSm },
2470 { X86::SETOr, X86::SETOm },
2471 { X86::SETPr, X86::SETPm },
2472 { X86::SETSr, X86::SETSm }
2473 };
2474
2475 assert(CC < 16 && "Can only handle standard cond codes");
2476 return Opc[CC][HasMemoryOperand ? 1 : 0];
2477}
2478
2479/// getCMovFromCond - Return a cmov opcode for the given condition,
2480/// register size in bytes, and operand type.
2481static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2482 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2483 static const uint16_t Opc[32][3] = {
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2484 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2485 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2486 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2487 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2488 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2489 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2490 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2491 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2492 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2493 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2494 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2495 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2496 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2497 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2498 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2499 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2500 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2501 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2502 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2503 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2504 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2505 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2506 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2507 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2508 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2509 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2510 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2511 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2512 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2513 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2514 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2515 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2516 };
2517
2518 assert(CC < 16 && "Can only handle standard cond codes");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2519 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2520 switch(RegBytes) {
2521 default: llvm_unreachable("Illegal register size!");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2522 case 2: return Opc[Idx][0];
2523 case 4: return Opc[Idx][1];
2524 case 8: return Opc[Idx][2];
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2525 }
2526}
2527
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2528bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2529 if (!MI->isTerminator()) return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2530
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2531 // Conditional branch is a special case.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2532 if (MI->isBranch() && !MI->isBarrier())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2533 return true;
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2534 if (!MI->isPredicable())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2535 return true;
2536 return !isPredicated(MI);
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2537}
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2538
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2539bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2540 MachineBasicBlock *&TBB,
2541 MachineBasicBlock *&FBB,
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2542 SmallVectorImpl<MachineOperand> &Cond,
2543 bool AllowModify) const {
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2544 // Start from the bottom of the block and work up, examining the
2545 // terminator instructions.
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2546 MachineBasicBlock::iterator I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2547 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2548 while (I != MBB.begin()) {
2549 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2550 if (I->isDebugValue())
2551 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2552
2553 // Working from the bottom, when we see a non-terminator instruction, we're
2554 // done.
Jakob Stoklund Olesenc30b4dd2010-07-16 17:41:44 +0000[diff] [blame]2555 if (!isUnpredicatedTerminator(I))
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2556 break;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2557
2558 // A terminator that isn't a branch can't easily be handled by this
2559 // analysis.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2560 if (!I->isBranch())
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2561 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2562
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2563 // Handle unconditional branches.
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2564 if (I->getOpcode() == X86::JMP_4) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2565 UnCondBrIter = I;
2566
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2567 if (!AllowModify) {
2568 TBB = I->getOperand(0).getMBB();
Evan Cheng2fa28112009-05-08 06:34:09 +0000[diff] [blame]2569 continue;
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2570 }
2571
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2572 // If the block has any instructions after a JMP, delete them.
Chris Lattnera48f44d2009-12-03 00:50:42 +0000[diff] [blame]2573 while (llvm::next(I) != MBB.end())
2574 llvm::next(I)->eraseFromParent();
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2575
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2576 Cond.clear();
2577 FBB = 0;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2578
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2579 // Delete the JMP if it's equivalent to a fall-through.
2580 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2581 TBB = 0;
2582 I->eraseFromParent();
2583 I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2584 UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2585 continue;
2586 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2587
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2588 // TBB is used to indicate the unconditional destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2589 TBB = I->getOperand(0).getMBB();
2590 continue;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2591 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2592
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2593 // Handle conditional branches.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2594 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2595 if (BranchCode == X86::COND_INVALID)
2596 return true; // Can't handle indirect branch.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2597
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2598 // Working from the bottom, handle the first conditional branch.
2599 if (Cond.empty()) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2600 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2601 if (AllowModify && UnCondBrIter != MBB.end() &&
2602 MBB.isLayoutSuccessor(TargetBB)) {
2603 // If we can modify the code and it ends in something like:
2604 //
2605 // jCC L1
2606 // jmp L2
2607 // L1:
2608 // ...
2609 // L2:
2610 //
2611 // Then we can change this to:
2612 //
2613 // jnCC L2
2614 // L1:
2615 // ...
2616 // L2:
2617 //
2618 // Which is a bit more efficient.
2619 // We conditionally jump to the fall-through block.
2620 BranchCode = GetOppositeBranchCondition(BranchCode);
2621 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2622 MachineBasicBlock::iterator OldInst = I;
2623
2624 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2625 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2626 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2627 .addMBB(TargetBB);
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2628
2629 OldInst->eraseFromParent();
2630 UnCondBrIter->eraseFromParent();
2631
2632 // Restart the analysis.
2633 UnCondBrIter = MBB.end();
2634 I = MBB.end();
2635 continue;
2636 }
2637
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2638 FBB = TBB;
2639 TBB = I->getOperand(0).getMBB();
2640 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2641 continue;
2642 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2643
2644 // Handle subsequent conditional branches. Only handle the case where all
2645 // conditional branches branch to the same destination and their condition
2646 // opcodes fit one of the special multi-branch idioms.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2647 assert(Cond.size() == 1);
2648 assert(TBB);
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2649
2650 // Only handle the case where all conditional branches branch to the same
2651 // destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2652 if (TBB != I->getOperand(0).getMBB())
2653 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2654
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2655 // If the conditions are the same, we can leave them alone.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2656 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2657 if (OldBranchCode == BranchCode)
2658 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2659
2660 // If they differ, see if they fit one of the known patterns. Theoretically,
2661 // we could handle more patterns here, but we shouldn't expect to see them
2662 // if instruction selection has done a reasonable job.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2663 if ((OldBranchCode == X86::COND_NP &&
2664 BranchCode == X86::COND_E) ||
2665 (OldBranchCode == X86::COND_E &&
2666 BranchCode == X86::COND_NP))
2667 BranchCode = X86::COND_NP_OR_E;
2668 else if ((OldBranchCode == X86::COND_P &&
2669 BranchCode == X86::COND_NE) ||
2670 (OldBranchCode == X86::COND_NE &&
2671 BranchCode == X86::COND_P))
2672 BranchCode = X86::COND_NE_OR_P;
2673 else
2674 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2675
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2676 // Update the MachineOperand.
2677 Cond[0].setImm(BranchCode);
Chris Lattner74436002006-10-30 22:27:23 +0000[diff] [blame]2678 }
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2679
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2680 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2681}
2682
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2683unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2684 MachineBasicBlock::iterator I = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2685 unsigned Count = 0;
2686
2687 while (I != MBB.begin()) {
2688 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2689 if (I->isDebugValue())
2690 continue;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2691 if (I->getOpcode() != X86::JMP_4 &&
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2692 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2693 break;
2694 // Remove the branch.
2695 I->eraseFromParent();
2696 I = MBB.end();
2697 ++Count;
2698 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2699
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2700 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2701}
2702
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2703unsigned
2704X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2705 MachineBasicBlock *FBB,
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2706 const SmallVectorImpl<MachineOperand> &Cond,
2707 DebugLoc DL) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2708 // Shouldn't be a fall through.
2709 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
Chris Lattner6fca75e2006-10-21 05:34:23 +0000[diff] [blame]2710 assert((Cond.size() == 1 || Cond.size() == 0) &&
2711 "X86 branch conditions have one component!");
2712
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2713 if (Cond.empty()) {
2714 // Unconditional branch?
2715 assert(!FBB && "Unconditional branch with multiple successors!");
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2716 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2717 return 1;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2718 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2719
2720 // Conditional branch.
2721 unsigned Count = 0;
2722 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2723 switch (CC) {
2724 case X86::COND_NP_OR_E:
2725 // Synthesize NP_OR_E with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2726 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2727 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2728 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2729 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2730 break;
2731 case X86::COND_NE_OR_P:
2732 // Synthesize NE_OR_P with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2733 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2734 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2735 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2736 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2737 break;
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2738 default: {
2739 unsigned Opc = GetCondBranchFromCond(CC);
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2740 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2741 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2742 }
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2743 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2744 if (FBB) {
2745 // Two-way Conditional branch. Insert the second branch.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2746 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2747 ++Count;
2748 }
2749 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2750}
2751
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2752bool X86InstrInfo::
2753canInsertSelect(const MachineBasicBlock &MBB,
2754 const SmallVectorImpl<MachineOperand> &Cond,
2755 unsigned TrueReg, unsigned FalseReg,
2756 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2757 // Not all subtargets have cmov instructions.
2758 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2759 return false;
2760 if (Cond.size() != 1)
2761 return false;
2762 // We cannot do the composite conditions, at least not in SSA form.
2763 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2764 return false;
2765
2766 // Check register classes.
2767 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2768 const TargetRegisterClass *RC =
2769 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2770 if (!RC)
2771 return false;
2772
2773 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2774 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2775 X86::GR32RegClass.hasSubClassEq(RC) ||
2776 X86::GR64RegClass.hasSubClassEq(RC)) {
2777 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2778 // Bridge. Probably Ivy Bridge as well.
2779 CondCycles = 2;
2780 TrueCycles = 2;
2781 FalseCycles = 2;
2782 return true;
2783 }
2784
2785 // Can't do vectors.
2786 return false;
2787}
2788
2789void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2790 MachineBasicBlock::iterator I, DebugLoc DL,
2791 unsigned DstReg,
2792 const SmallVectorImpl<MachineOperand> &Cond,
2793 unsigned TrueReg, unsigned FalseReg) const {
2794 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2795 assert(Cond.size() == 1 && "Invalid Cond array");
2796 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2797 MRI.getRegClass(DstReg)->getSize(),
2798 false/*HasMemoryOperand*/);
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2799 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2800}
2801
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2802/// isHReg - Test if the given register is a physical h register.
2803static bool isHReg(unsigned Reg) {
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2804 return X86::GR8_ABCD_HRegClass.contains(Reg);
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2805}
2806
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2807// Try and copy between VR128/VR64 and GR64 registers.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2808static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2809 bool HasAVX) {
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2810 // SrcReg(VR128) -> DestReg(GR64)
2811 // SrcReg(VR64) -> DestReg(GR64)
2812 // SrcReg(GR64) -> DestReg(VR128)
2813 // SrcReg(GR64) -> DestReg(VR64)
2814
2815 if (X86::GR64RegClass.contains(DestReg)) {
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2816 if (X86::VR128RegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2817 // Copy from a VR128 register to a GR64 register.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2818 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2819 if (X86::VR64RegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2820 // Copy from a VR64 register to a GR64 register.
2821 return X86::MOVSDto64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2822 } else if (X86::GR64RegClass.contains(SrcReg)) {
2823 // Copy from a GR64 register to a VR128 register.
2824 if (X86::VR128RegClass.contains(DestReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2825 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2826 // Copy from a GR64 register to a VR64 register.
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2827 if (X86::VR64RegClass.contains(DestReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2828 return X86::MOV64toSDrr;
2829 }
2830
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2831 // SrcReg(FR32) -> DestReg(GR32)
2832 // SrcReg(GR32) -> DestReg(FR32)
2833
2834 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2835 // Copy from a FR32 register to a GR32 register.
2836 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2837
2838 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2839 // Copy from a GR32 register to a FR32 register.
2840 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2841
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2842 return 0;
2843}
2844
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2845void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2846 MachineBasicBlock::iterator MI, DebugLoc DL,
2847 unsigned DestReg, unsigned SrcReg,
2848 bool KillSrc) const {
2849 // First deal with the normal symmetric copies.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2850 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2851 unsigned Opc;
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2852 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2853 Opc = X86::MOV64rr;
2854 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2855 Opc = X86::MOV32rr;
2856 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2857 Opc = X86::MOV16rr;
2858 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2859 // Copying to or from a physical H register on x86-64 requires a NOREX
2860 // move. Otherwise use a normal move.
2861 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2862 TM.getSubtarget<X86Subtarget>().is64Bit()) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2863 Opc = X86::MOV8rr_NOREX;
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2864 // Both operands must be encodable without an REX prefix.
2865 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2866 "8-bit H register can not be copied outside GR8_NOREX");
2867 } else
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2868 Opc = X86::MOV8rr;
2869 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2870 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]2871 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2872 Opc = X86::VMOVAPSYrr;
Jakob Stoklund Olesenec58a432010-07-08 22:30:35 +0000[diff] [blame]2873 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2874 Opc = X86::MMX_MOVQ64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2875 else
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2876 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2877
2878 if (Opc) {
2879 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2880 .addReg(SrcReg, getKillRegState(KillSrc));
2881 return;
2882 }
2883
2884 // Moving EFLAGS to / from another register requires a push and a pop.
Nadav Rotemd5aae982012-12-21 23:48:49 +0000[diff] [blame]2885 // Notice that we have to adjust the stack if we don't want to clobber the
2886 // first frame index. See X86FrameLowering.cpp - colobbersTheStack.
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2887 if (SrcReg == X86::EFLAGS) {
2888 if (X86::GR64RegClass.contains(DestReg)) {
2889 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2890 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2891 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2892 }
2893 if (X86::GR32RegClass.contains(DestReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2894 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2895 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2896 return;
2897 }
2898 }
2899 if (DestReg == X86::EFLAGS) {
2900 if (X86::GR64RegClass.contains(SrcReg)) {
2901 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2902 .addReg(SrcReg, getKillRegState(KillSrc));
2903 BuildMI(MBB, MI, DL, get(X86::POPF64));
2904 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2905 }
2906 if (X86::GR32RegClass.contains(SrcReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2907 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2908 .addReg(SrcReg, getKillRegState(KillSrc));
2909 BuildMI(MBB, MI, DL, get(X86::POPF32));
2910 return;
2911 }
2912 }
2913
2914 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2915 << " to " << RI.getName(DestReg) << '\n');
2916 llvm_unreachable("Cannot emit physreg copy instruction");
2917}
2918
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2919static unsigned getLoadStoreRegOpcode(unsigned Reg,
2920 const TargetRegisterClass *RC,
2921 bool isStackAligned,
2922 const TargetMachine &TM,
2923 bool load) {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2924 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2925 switch (RC->getSize()) {
Rafael Espindola6635f982010-07-12 03:43:04 +0000[diff] [blame]2926 default:
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2927 llvm_unreachable("Unknown spill size");
2928 case 1:
2929 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2930 if (TM.getSubtarget<X86Subtarget>().is64Bit())
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2931 // Copying to or from a physical H register on x86-64 requires a NOREX
2932 // move. Otherwise use a normal move.
2933 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2934 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2935 return load ? X86::MOV8rm : X86::MOV8mr;
2936 case 2:
2937 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2938 return load ? X86::MOV16rm : X86::MOV16mr;
2939 case 4:
2940 if (X86::GR32RegClass.hasSubClassEq(RC))
2941 return load ? X86::MOV32rm : X86::MOV32mr;
2942 if (X86::FR32RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2943 return load ?
2944 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2945 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2946 if (X86::RFP32RegClass.hasSubClassEq(RC))
2947 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2948 llvm_unreachable("Unknown 4-byte regclass");
2949 case 8:
2950 if (X86::GR64RegClass.hasSubClassEq(RC))
2951 return load ? X86::MOV64rm : X86::MOV64mr;
2952 if (X86::FR64RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2953 return load ?
2954 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
2955 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2956 if (X86::VR64RegClass.hasSubClassEq(RC))
2957 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2958 if (X86::RFP64RegClass.hasSubClassEq(RC))
2959 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2960 llvm_unreachable("Unknown 8-byte regclass");
2961 case 10:
2962 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2963 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2964 case 16: {
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2965 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2966 // If stack is realigned we can use aligned stores.
2967 if (isStackAligned)
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2968 return load ?
2969 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
2970 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2971 else
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2972 return load ?
2973 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
2974 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
2975 }
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]2976 case 32:
2977 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2978 // If stack is realigned we can use aligned stores.
2979 if (isStackAligned)
2980 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2981 else
2982 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2983 }
2984}
2985
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2986static unsigned getStoreRegOpcode(unsigned SrcReg,
2987 const TargetRegisterClass *RC,
2988 bool isStackAligned,
2989 TargetMachine &TM) {
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2990 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2991}
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2992
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2993
2994static unsigned getLoadRegOpcode(unsigned DestReg,
2995 const TargetRegisterClass *RC,
2996 bool isStackAligned,
2997 const TargetMachine &TM) {
2998 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2999}
3000
3001void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3002 MachineBasicBlock::iterator MI,
3003 unsigned SrcReg, bool isKill, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]3004 const TargetRegisterClass *RC,
3005 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3006 const MachineFunction &MF = *MBB.getParent();
Jakob Stoklund Olesenc3c05ed2010-07-27 04:16:58 +0000[diff] [blame]3007 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
3008 "Stack slot too small for store");
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3009 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3010 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]3011 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3012 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]3013 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3014 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]3015 .addReg(SrcReg, getKillRegState(isKill));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3016}
3017
3018void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
3019 bool isKill,
3020 SmallVectorImpl<MachineOperand> &Addr,
3021 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3022 MachineInstr::mmo_iterator MMOBegin,
3023 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3024 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3025 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3026 bool isAligned = MMOBegin != MMOEnd &&
3027 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3028 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]3029 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]3030 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3031 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3032 MIB.addOperand(Addr[i]);
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]3033 MIB.addReg(SrcReg, getKillRegState(isKill));
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3034 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3035 NewMIs.push_back(MIB);
3036}
3037
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3038
3039void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3040 MachineBasicBlock::iterator MI,
3041 unsigned DestReg, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]3042 const TargetRegisterClass *RC,
3043 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]3044 const MachineFunction &MF = *MBB.getParent();
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3045 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3046 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]3047 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3048 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]3049 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3050 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3051}
3052
3053void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]3054 SmallVectorImpl<MachineOperand> &Addr,
3055 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3056 MachineInstr::mmo_iterator MMOBegin,
3057 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3058 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3059 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3060 bool isAligned = MMOBegin != MMOEnd &&
3061 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3062 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]3063 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]3064 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3065 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3066 MIB.addOperand(Addr[i]);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3067 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3068 NewMIs.push_back(MIB);
3069}
3070
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3071bool X86InstrInfo::
3072analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3073 int &CmpMask, int &CmpValue) const {
3074 switch (MI->getOpcode()) {
3075 default: break;
3076 case X86::CMP64ri32:
3077 case X86::CMP64ri8:
3078 case X86::CMP32ri:
3079 case X86::CMP32ri8:
3080 case X86::CMP16ri:
3081 case X86::CMP16ri8:
3082 case X86::CMP8ri:
3083 SrcReg = MI->getOperand(0).getReg();
3084 SrcReg2 = 0;
3085 CmpMask = ~0;
3086 CmpValue = MI->getOperand(1).getImm();
3087 return true;
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3088 // A SUB can be used to perform comparison.
3089 case X86::SUB64rm:
3090 case X86::SUB32rm:
3091 case X86::SUB16rm:
3092 case X86::SUB8rm:
3093 SrcReg = MI->getOperand(1).getReg();
3094 SrcReg2 = 0;
3095 CmpMask = ~0;
3096 CmpValue = 0;
3097 return true;
3098 case X86::SUB64rr:
3099 case X86::SUB32rr:
3100 case X86::SUB16rr:
3101 case X86::SUB8rr:
3102 SrcReg = MI->getOperand(1).getReg();
3103 SrcReg2 = MI->getOperand(2).getReg();
3104 CmpMask = ~0;
3105 CmpValue = 0;
3106 return true;
3107 case X86::SUB64ri32:
3108 case X86::SUB64ri8:
3109 case X86::SUB32ri:
3110 case X86::SUB32ri8:
3111 case X86::SUB16ri:
3112 case X86::SUB16ri8:
3113 case X86::SUB8ri:
3114 SrcReg = MI->getOperand(1).getReg();
3115 SrcReg2 = 0;
3116 CmpMask = ~0;
3117 CmpValue = MI->getOperand(2).getImm();
3118 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3119 case X86::CMP64rr:
3120 case X86::CMP32rr:
3121 case X86::CMP16rr:
3122 case X86::CMP8rr:
3123 SrcReg = MI->getOperand(0).getReg();
3124 SrcReg2 = MI->getOperand(1).getReg();
3125 CmpMask = ~0;
3126 CmpValue = 0;
3127 return true;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3128 case X86::TEST8rr:
3129 case X86::TEST16rr:
3130 case X86::TEST32rr:
3131 case X86::TEST64rr:
3132 SrcReg = MI->getOperand(0).getReg();
3133 if (MI->getOperand(1).getReg() != SrcReg) return false;
3134 // Compare against zero.
3135 SrcReg2 = 0;
3136 CmpMask = ~0;
3137 CmpValue = 0;
3138 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3139 }
3140 return false;
3141}
3142
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3143/// isRedundantFlagInstr - check whether the first instruction, whose only
3144/// purpose is to update flags, can be made redundant.
3145/// CMPrr can be made redundant by SUBrr if the operands are the same.
3146/// This function can be extended later on.
3147/// SrcReg, SrcRegs: register operands for FlagI.
3148/// ImmValue: immediate for FlagI if it takes an immediate.
3149inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3150 unsigned SrcReg2, int ImmValue,
3151 MachineInstr *OI) {
3152 if (((FlagI->getOpcode() == X86::CMP64rr &&
3153 OI->getOpcode() == X86::SUB64rr) ||
3154 (FlagI->getOpcode() == X86::CMP32rr &&
3155 OI->getOpcode() == X86::SUB32rr)||
3156 (FlagI->getOpcode() == X86::CMP16rr &&
3157 OI->getOpcode() == X86::SUB16rr)||
3158 (FlagI->getOpcode() == X86::CMP8rr &&
3159 OI->getOpcode() == X86::SUB8rr)) &&
3160 ((OI->getOperand(1).getReg() == SrcReg &&
3161 OI->getOperand(2).getReg() == SrcReg2) ||
3162 (OI->getOperand(1).getReg() == SrcReg2 &&
3163 OI->getOperand(2).getReg() == SrcReg)))
3164 return true;
3165
3166 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3167 OI->getOpcode() == X86::SUB64ri32) ||
3168 (FlagI->getOpcode() == X86::CMP64ri8 &&
3169 OI->getOpcode() == X86::SUB64ri8) ||
3170 (FlagI->getOpcode() == X86::CMP32ri &&
3171 OI->getOpcode() == X86::SUB32ri) ||
3172 (FlagI->getOpcode() == X86::CMP32ri8 &&
3173 OI->getOpcode() == X86::SUB32ri8) ||
3174 (FlagI->getOpcode() == X86::CMP16ri &&
3175 OI->getOpcode() == X86::SUB16ri) ||
3176 (FlagI->getOpcode() == X86::CMP16ri8 &&
3177 OI->getOpcode() == X86::SUB16ri8) ||
3178 (FlagI->getOpcode() == X86::CMP8ri &&
3179 OI->getOpcode() == X86::SUB8ri)) &&
3180 OI->getOperand(1).getReg() == SrcReg &&
3181 OI->getOperand(2).getImm() == ImmValue)
3182 return true;
3183 return false;
3184}
3185
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3186/// isDefConvertible - check whether the definition can be converted
3187/// to remove a comparison against zero.
3188inline static bool isDefConvertible(MachineInstr *MI) {
3189 switch (MI->getOpcode()) {
3190 default: return false;
David Majnemer7ea2a522013-05-22 08:13:02 +0000[diff] [blame^]3191
3192 // The shift instructions only modify ZF if their shift count is non-zero.
3193 // N.B.: The processor truncates the shift count depending on the encoding.
3194 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3195 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3196 return getTruncatedShiftCount(MI, 2) != 0;
3197
3198 // Some left shift instructions can be turned into LEA instructions but only
3199 // if their flags aren't used. Avoid transforming such instructions.
3200 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3201 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3202 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3203 return ShAmt != 0;
3204 }
3205
3206 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3207 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3208 return getTruncatedShiftCount(MI, 3) != 0;
3209
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3210 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3211 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3212 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3213 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3214 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
Craig Topper5b08cf72012-12-17 04:55:07 +0000[diff] [blame]3215 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3216 case X86::DEC64_32r: case X86::DEC64_16r:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3217 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3218 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3219 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3220 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3221 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
Craig Topper5b08cf72012-12-17 04:55:07 +0000[diff] [blame]3222 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3223 case X86::INC64_32r: case X86::INC64_16r:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3224 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3225 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3226 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3227 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3228 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3229 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3230 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3231 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3232 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3233 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3234 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3235 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3236 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3237 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3238 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3239 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3240 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3241 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3242 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3243 case X86::ADC32ri: case X86::ADC32ri8:
3244 case X86::ADC32rr: case X86::ADC64ri32:
3245 case X86::ADC64ri8: case X86::ADC64rr:
3246 case X86::SBB32ri: case X86::SBB32ri8:
3247 case X86::SBB32rr: case X86::SBB64ri32:
3248 case X86::SBB64ri8: case X86::SBB64rr:
Craig Topperf3ff6ae2012-12-17 05:12:30 +0000[diff] [blame]3249 case X86::ANDN32rr: case X86::ANDN32rm:
3250 case X86::ANDN64rr: case X86::ANDN64rm:
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3251 case X86::BEXTR32rr: case X86::BEXTR64rr:
3252 case X86::BEXTR32rm: case X86::BEXTR64rm:
3253 case X86::BLSI32rr: case X86::BLSI32rm:
3254 case X86::BLSI64rr: case X86::BLSI64rm:
3255 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3256 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3257 case X86::BLSR32rr: case X86::BLSR32rm:
3258 case X86::BLSR64rr: case X86::BLSR64rm:
3259 case X86::BZHI32rr: case X86::BZHI32rm:
3260 case X86::BZHI64rr: case X86::BZHI64rm:
3261 case X86::LZCNT16rr: case X86::LZCNT16rm:
3262 case X86::LZCNT32rr: case X86::LZCNT32rm:
3263 case X86::LZCNT64rr: case X86::LZCNT64rm:
3264 case X86::POPCNT16rr:case X86::POPCNT16rm:
3265 case X86::POPCNT32rr:case X86::POPCNT32rm:
3266 case X86::POPCNT64rr:case X86::POPCNT64rm:
3267 case X86::TZCNT16rr: case X86::TZCNT16rm:
3268 case X86::TZCNT32rr: case X86::TZCNT32rm:
3269 case X86::TZCNT64rr: case X86::TZCNT64rm:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3270 return true;
3271 }
3272}
3273
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3274/// optimizeCompareInstr - Check if there exists an earlier instruction that
3275/// operates on the same source operands and sets flags in the same way as
3276/// Compare; remove Compare if possible.
3277bool X86InstrInfo::
3278optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3279 int CmpMask, int CmpValue,
3280 const MachineRegisterInfo *MRI) const {
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3281 // Check whether we can replace SUB with CMP.
3282 unsigned NewOpcode = 0;
3283 switch (CmpInstr->getOpcode()) {
3284 default: break;
3285 case X86::SUB64ri32:
3286 case X86::SUB64ri8:
3287 case X86::SUB32ri:
3288 case X86::SUB32ri8:
3289 case X86::SUB16ri:
3290 case X86::SUB16ri8:
3291 case X86::SUB8ri:
3292 case X86::SUB64rm:
3293 case X86::SUB32rm:
3294 case X86::SUB16rm:
3295 case X86::SUB8rm:
3296 case X86::SUB64rr:
3297 case X86::SUB32rr:
3298 case X86::SUB16rr:
3299 case X86::SUB8rr: {
3300 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3301 return false;
3302 // There is no use of the destination register, we can replace SUB with CMP.
3303 switch (CmpInstr->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]3304 default: llvm_unreachable("Unreachable!");
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3305 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3306 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3307 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3308 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3309 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3310 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3311 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3312 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3313 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3314 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3315 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3316 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3317 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3318 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3319 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3320 }
3321 CmpInstr->setDesc(get(NewOpcode));
3322 CmpInstr->RemoveOperand(0);
3323 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3324 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3325 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3326 return false;
3327 }
3328 }
3329
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3330 // Get the unique definition of SrcReg.
3331 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3332 if (!MI) return false;
3333
3334 // CmpInstr is the first instruction of the BB.
3335 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3336
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3337 // If we are comparing against zero, check whether we can use MI to update
3338 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3339 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3340 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3341 !isDefConvertible(MI)))
3342 return false;
3343
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3344 // We are searching for an earlier instruction that can make CmpInstr
3345 // redundant and that instruction will be saved in Sub.
3346 MachineInstr *Sub = NULL;
3347 const TargetRegisterInfo *TRI = &getRegisterInfo();
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3348
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3349 // We iterate backward, starting from the instruction before CmpInstr and
3350 // stop when reaching the definition of a source register or done with the BB.
3351 // RI points to the instruction before CmpInstr.
3352 // If the definition is in this basic block, RE points to the definition;
3353 // otherwise, RE is the rend of the basic block.
3354 MachineBasicBlock::reverse_iterator
3355 RI = MachineBasicBlock::reverse_iterator(I),
3356 RE = CmpInstr->getParent() == MI->getParent() ?
3357 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3358 CmpInstr->getParent()->rend();
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3359 MachineInstr *Movr0Inst = 0;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3360 for (; RI != RE; ++RI) {
3361 MachineInstr *Instr = &*RI;
3362 // Check whether CmpInstr can be made redundant by the current instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3363 if (!IsCmpZero &&
3364 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3365 Sub = Instr;
3366 break;
3367 }
3368
3369 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3370 Instr->readsRegister(X86::EFLAGS, TRI)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3371 // This instruction modifies or uses EFLAGS.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3372
3373 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3374 // They are safe to move up, if the definition to EFLAGS is dead and
3375 // earlier instructions do not read or write EFLAGS.
3376 if (!Movr0Inst && (Instr->getOpcode() == X86::MOV8r0 ||
3377 Instr->getOpcode() == X86::MOV16r0 ||
3378 Instr->getOpcode() == X86::MOV32r0 ||
3379 Instr->getOpcode() == X86::MOV64r0) &&
3380 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3381 Movr0Inst = Instr;
3382 continue;
3383 }
3384
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3385 // We can't remove CmpInstr.
3386 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3387 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3388 }
3389
3390 // Return false if no candidates exist.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3391 if (!IsCmpZero && !Sub)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3392 return false;
3393
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3394 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3395 Sub->getOperand(2).getReg() == SrcReg);
3396
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3397 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3398 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3399 // If we are done with the basic block, we need to check whether EFLAGS is
3400 // live-out.
3401 bool IsSafe = false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3402 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3403 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3404 for (++I; I != E; ++I) {
3405 const MachineInstr &Instr = *I;
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3406 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3407 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3408 // We should check the usage if this instruction uses and updates EFLAGS.
3409 if (!UseEFLAGS && ModifyEFLAGS) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3410 // It is safe to remove CmpInstr if EFLAGS is updated again.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3411 IsSafe = true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3412 break;
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3413 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3414 if (!UseEFLAGS && !ModifyEFLAGS)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3415 continue;
3416
3417 // EFLAGS is used by this instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3418 X86::CondCode OldCC;
3419 bool OpcIsSET = false;
3420 if (IsCmpZero || IsSwapped) {
3421 // We decode the condition code from opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3422 if (Instr.isBranch())
3423 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3424 else {
3425 OldCC = getCondFromSETOpc(Instr.getOpcode());
3426 if (OldCC != X86::COND_INVALID)
3427 OpcIsSET = true;
3428 else
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]3429 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3430 }
3431 if (OldCC == X86::COND_INVALID) return false;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3432 }
3433 if (IsCmpZero) {
3434 switch (OldCC) {
3435 default: break;
3436 case X86::COND_A: case X86::COND_AE:
3437 case X86::COND_B: case X86::COND_BE:
3438 case X86::COND_G: case X86::COND_GE:
3439 case X86::COND_L: case X86::COND_LE:
3440 case X86::COND_O: case X86::COND_NO:
3441 // CF and OF are used, we can't perform this optimization.
3442 return false;
3443 }
3444 } else if (IsSwapped) {
3445 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3446 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3447 // We swap the condition code and synthesize the new opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3448 X86::CondCode NewCC = getSwappedCondition(OldCC);
3449 if (NewCC == X86::COND_INVALID) return false;
3450
3451 // Synthesize the new opcode.
3452 bool HasMemoryOperand = Instr.hasOneMemOperand();
3453 unsigned NewOpc;
3454 if (Instr.isBranch())
3455 NewOpc = GetCondBranchFromCond(NewCC);
3456 else if(OpcIsSET)
3457 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3458 else {
3459 unsigned DstReg = Instr.getOperand(0).getReg();
3460 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3461 HasMemoryOperand);
3462 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3463
3464 // Push the MachineInstr to OpsToUpdate.
3465 // If it is safe to remove CmpInstr, the condition code of these
3466 // instructions will be modified.
3467 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3468 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3469 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3470 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3471 IsSafe = true;
3472 break;
3473 }
3474 }
3475
3476 // If EFLAGS is not killed nor re-defined, we should check whether it is
3477 // live-out. If it is live-out, do not optimize.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3478 if ((IsCmpZero || IsSwapped) && !IsSafe) {
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3479 MachineBasicBlock *MBB = CmpInstr->getParent();
3480 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3481 SE = MBB->succ_end(); SI != SE; ++SI)
3482 if ((*SI)->isLiveIn(X86::EFLAGS))
3483 return false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3484 }
3485
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3486 // The instruction to be updated is either Sub or MI.
3487 Sub = IsCmpZero ? MI : Sub;
David Majnemer5ba473a2013-05-18 01:02:03 +0000[diff] [blame]3488 // Move Movr0Inst to the appropriate place before Sub.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3489 if (Movr0Inst) {
David Majnemer5ba473a2013-05-18 01:02:03 +0000[diff] [blame]3490 // Look backwards until we find a def that doesn't use the current EFLAGS.
3491 Def = Sub;
3492 MachineBasicBlock::reverse_iterator
3493 InsertI = MachineBasicBlock::reverse_iterator(++Def),
3494 InsertE = Sub->getParent()->rend();
3495 for (; InsertI != InsertE; ++InsertI) {
3496 MachineInstr *Instr = &*InsertI;
3497 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3498 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3499 Sub->getParent()->remove(Movr0Inst);
3500 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3501 Movr0Inst);
3502 break;
3503 }
3504 }
3505 if (InsertI == InsertE)
3506 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3507 }
3508
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3509 // Make sure Sub instruction defines EFLAGS and mark the def live.
David Majnemer8f169742013-05-15 22:03:08 +0000[diff] [blame]3510 unsigned i = 0, e = Sub->getNumOperands();
3511 for (; i != e; ++i) {
3512 MachineOperand &MO = Sub->getOperand(i);
3513 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3514 MO.setIsDead(false);
3515 break;
3516 }
3517 }
3518 assert(i != e && "Unable to locate a def EFLAGS operand");
3519
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3520 CmpInstr->eraseFromParent();
3521
3522 // Modify the condition code of instructions in OpsToUpdate.
3523 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3524 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3525 return true;
3526}
3527
Manman Ren5759d012012-08-02 00:56:42 +0000[diff] [blame]3528/// optimizeLoadInstr - Try to remove the load by folding it to a register
3529/// operand at the use. We fold the load instructions if load defines a virtual
3530/// register, the virtual register is used once in the same BB, and the
3531/// instructions in-between do not load or store, and have no side effects.
3532MachineInstr* X86InstrInfo::
3533optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3534 unsigned &FoldAsLoadDefReg,
3535 MachineInstr *&DefMI) const {
3536 if (FoldAsLoadDefReg == 0)
3537 return 0;
3538 // To be conservative, if there exists another load, clear the load candidate.
3539 if (MI->mayLoad()) {
3540 FoldAsLoadDefReg = 0;
3541 return 0;
3542 }
3543
3544 // Check whether we can move DefMI here.
3545 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3546 assert(DefMI);
3547 bool SawStore = false;
3548 if (!DefMI->isSafeToMove(this, 0, SawStore))
3549 return 0;
3550
3551 // We try to commute MI if possible.
3552 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3553 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3554 // Collect information about virtual register operands of MI.
3555 unsigned SrcOperandId = 0;
3556 bool FoundSrcOperand = false;
3557 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3558 MachineOperand &MO = MI->getOperand(i);
3559 if (!MO.isReg())
3560 continue;
3561 unsigned Reg = MO.getReg();
3562 if (Reg != FoldAsLoadDefReg)
3563 continue;
3564 // Do not fold if we have a subreg use or a def or multiple uses.
3565 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3566 return 0;
3567
3568 SrcOperandId = i;
3569 FoundSrcOperand = true;
3570 }
3571 if (!FoundSrcOperand) return 0;
3572
3573 // Check whether we can fold the def into SrcOperandId.
3574 SmallVector<unsigned, 8> Ops;
3575 Ops.push_back(SrcOperandId);
3576 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3577 if (FoldMI) {
3578 FoldAsLoadDefReg = 0;
3579 return FoldMI;
3580 }
3581
3582 if (Idx == 1) {
3583 // MI was changed but it didn't help, commute it back!
3584 commuteInstruction(MI, false);
3585 return 0;
3586 }
3587
3588 // Check whether we can commute MI and enable folding.
3589 if (MI->isCommutable()) {
3590 MachineInstr *NewMI = commuteInstruction(MI, false);
3591 // Unable to commute.
3592 if (!NewMI) return 0;
3593 if (NewMI != MI) {
3594 // New instruction. It doesn't need to be kept.
3595 NewMI->eraseFromParent();
3596 return 0;
3597 }
3598 }
3599 }
3600 return 0;
3601}
3602
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3603/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3604/// instruction with two undef reads of the register being defined. This is
3605/// used for mapping:
3606/// %xmm4 = V_SET0
3607/// to:
3608/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3609///
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3610static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3611 const MCInstrDesc &Desc) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3612 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3613 unsigned Reg = MIB->getOperand(0).getReg();
3614 MIB->setDesc(Desc);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3615
3616 // MachineInstr::addOperand() will insert explicit operands before any
3617 // implicit operands.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3618 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3619 // But we don't trust that.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3620 assert(MIB->getOperand(1).getReg() == Reg &&
3621 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3622 return true;
3623}
3624
3625bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3626 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3627 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3628 switch (MI->getOpcode()) {
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3629 case X86::SETB_C8r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3630 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3631 case X86::SETB_C16r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3632 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3633 case X86::SETB_C32r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3634 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
Craig Topper93849022012-10-05 06:05:15 +0000[diff] [blame]3635 case X86::SETB_C64r:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3636 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3637 case X86::V_SET0:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3638 case X86::FsFLD0SS:
3639 case X86::FsFLD0SD:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3640 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]3641 case X86::AVX_SET0:
3642 assert(HasAVX && "AVX not supported");
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3643 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3644 case X86::V_SETALLONES:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3645 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3646 case X86::AVX2_SETALLONES:
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3647 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
Jakob Stoklund Olesen729abd32011-10-08 18:28:28 +0000[diff] [blame]3648 case X86::TEST8ri_NOREX:
3649 MI->setDesc(get(X86::TEST8ri));
3650 return true;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3651 }
3652 return false;
3653}
3654
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3655MachineInstr*
3656X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
Evan Cheng250e9172010-04-29 01:13:30 +0000[diff] [blame]3657 int FrameIx, uint64_t Offset,
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3658 const MDNode *MDPtr,
3659 DebugLoc DL) const {
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3660 X86AddressMode AM;
3661 AM.BaseType = X86AddressMode::FrameIndexBase;
3662 AM.Base.FrameIndex = FrameIx;
3663 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
3664 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
3665 return &*MIB;
3666}
3667
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3668static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3669 const SmallVectorImpl<MachineOperand> &MOs,
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3670 MachineInstr *MI,
3671 const TargetInstrInfo &TII) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3672 // Create the base instruction with the memory operand as the first part.
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3673 // Omit the implicit operands, something BuildMI can't do.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3674 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3675 MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3676 MachineInstrBuilder MIB(MF, NewMI);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3677 unsigned NumAddrOps = MOs.size();
3678 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3679 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3680 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3681 addOffset(MIB, 0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3682
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3683 // Loop over the rest of the ri operands, converting them over.
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3684 unsigned NumOps = MI->getDesc().getNumOperands()-2;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3685 for (unsigned i = 0; i != NumOps; ++i) {
3686 MachineOperand &MO = MI->getOperand(i+2);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3687 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3688 }
3689 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3690 MachineOperand &MO = MI->getOperand(i);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3691 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3692 }
3693 return MIB;
3694}
3695
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3696static MachineInstr *FuseInst(MachineFunction &MF,
3697 unsigned Opcode, unsigned OpNo,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3698 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3699 MachineInstr *MI, const TargetInstrInfo &TII) {
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3700 // Omit the implicit operands, something BuildMI can't do.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3701 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3702 MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]3703 MachineInstrBuilder MIB(MF, NewMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3704
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3705 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3706 MachineOperand &MO = MI->getOperand(i);
3707 if (i == OpNo) {
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3708 assert(MO.isReg() && "Expected to fold into reg operand!");
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3709 unsigned NumAddrOps = MOs.size();
3710 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3711 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3712 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3713 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3714 } else {
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3715 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3716 }
3717 }
3718 return MIB;
3719}
3720
3721static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3722 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3723 MachineInstr *MI) {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3724 MachineFunction &MF = *MI->getParent()->getParent();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3725 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3726
3727 unsigned NumAddrOps = MOs.size();
3728 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3729 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3730 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3731 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3732 return MIB.addImm(0);
3733}
3734
3735MachineInstr*
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3736X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3737 MachineInstr *MI, unsigned i,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3738 const SmallVectorImpl<MachineOperand> &MOs,
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3739 unsigned Size, unsigned Align) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3740 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
Preston Gurdd6be4bf2013-03-27 23:16:18 +0000[diff] [blame]3741 bool isCallRegIndirect = TM.getSubtarget<X86Subtarget>().callRegIndirect();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3742 bool isTwoAddrFold = false;
Preston Gurdd6be4bf2013-03-27 23:16:18 +0000[diff] [blame]3743
3744 // Atom favors register form of call. So, we do not fold loads into calls
3745 // when X86Subtarget is Atom.
3746 if (isCallRegIndirect &&
3747 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
3748 return NULL;
3749 }
3750
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3751 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3752 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]3753 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3754
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]3755 // FIXME: AsmPrinter doesn't know how to handle
3756 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3757 if (MI->getOpcode() == X86::ADD32ri &&
3758 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3759 return NULL;
3760
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3761 MachineInstr *NewMI = NULL;
3762 // Folding a memory location into the two-address part of a two-address
3763 // instruction is different than folding it other places. It requires
3764 // replacing the *two* registers with the memory location.
3765 if (isTwoAddr && NumOps >= 2 && i < 2 &&
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3766 MI->getOperand(0).isReg() &&
3767 MI->getOperand(1).isReg() &&
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3768 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3769 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3770 isTwoAddrFold = true;
3771 } else if (i == 0) { // If operand 0
Craig Topperf9115972012-08-23 04:57:36 +0000[diff] [blame]3772 unsigned Opc = 0;
3773 switch (MI->getOpcode()) {
3774 default: break;
3775 case X86::MOV64r0: Opc = X86::MOV64mi32; break;
3776 case X86::MOV32r0: Opc = X86::MOV32mi; break;
3777 case X86::MOV16r0: Opc = X86::MOV16mi; break;
3778 case X86::MOV8r0: Opc = X86::MOV8mi; break;
3779 }
3780 if (Opc)
3781 NewMI = MakeM0Inst(*this, Opc, MOs, MI);
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]3782 if (NewMI)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3783 return NewMI;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3784
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3785 OpcodeTablePtr = &RegOp2MemOpTable0;
3786 } else if (i == 1) {
3787 OpcodeTablePtr = &RegOp2MemOpTable1;
3788 } else if (i == 2) {
3789 OpcodeTablePtr = &RegOp2MemOpTable2;
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]3790 } else if (i == 3) {
3791 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3792 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3793
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3794 // If table selected...
3795 if (OpcodeTablePtr) {
3796 // Find the Opcode to fuse
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3797 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3798 OpcodeTablePtr->find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3799 if (I != OpcodeTablePtr->end()) {
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3800 unsigned Opcode = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]3801 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3802 if (Align < MinAlign)
3803 return NULL;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3804 bool NarrowToMOV32rm = false;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3805 if (Size) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]3806 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3807 if (Size < RCSize) {
3808 // Check if it's safe to fold the load. If the size of the object is
3809 // narrower than the load width, then it's not.
3810 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3811 return NULL;
3812 // If this is a 64-bit load, but the spill slot is 32, then we can do
3813 // a 32-bit load which is implicitly zero-extended. This likely is due
3814 // to liveintervalanalysis remat'ing a load from stack slot.
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3815 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3816 return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3817 Opcode = X86::MOV32rm;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3818 NarrowToMOV32rm = true;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3819 }
3820 }
3821
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3822 if (isTwoAddrFold)
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3823 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3824 else
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3825 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3826
3827 if (NarrowToMOV32rm) {
3828 // If this is the special case where we use a MOV32rm to load a 32-bit
3829 // value and zero-extend the top bits. Change the destination register
3830 // to a 32-bit one.
3831 unsigned DstReg = NewMI->getOperand(0).getReg();
3832 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3833 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3834 X86::sub_32bit));
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3835 else
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3836 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3837 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3838 return NewMI;
3839 }
3840 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3841
3842 // No fusion
Jakob Stoklund Olesen51702ec2010-07-09 20:43:09 +0000[diff] [blame]3843 if (PrintFailedFusing && !MI->isCopy())
David Greened589daf2010-01-05 01:29:29 +0000[diff] [blame]3844 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3845 return NULL;
3846}
3847
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3848/// hasPartialRegUpdate - Return true for all instructions that only update
3849/// the first 32 or 64-bits of the destination register and leave the rest
3850/// unmodified. This can be used to avoid folding loads if the instructions
3851/// only update part of the destination register, and the non-updated part is
3852/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3853/// instructions breaks the partial register dependency and it can improve
3854/// performance. e.g.:
3855///
3856/// movss (%rdi), %xmm0
3857/// cvtss2sd %xmm0, %xmm0
3858///
3859/// Instead of
3860/// cvtss2sd (%rdi), %xmm0
3861///
Bruno Cardoso Lopes7b435682011-09-15 23:04:24 +0000[diff] [blame]3862/// FIXME: This should be turned into a TSFlags.
3863///
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3864static bool hasPartialRegUpdate(unsigned Opcode) {
3865 switch (Opcode) {
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]3866 case X86::CVTSI2SSrr:
3867 case X86::CVTSI2SS64rr:
3868 case X86::CVTSI2SDrr:
3869 case X86::CVTSI2SD64rr:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3870 case X86::CVTSD2SSrr:
3871 case X86::Int_CVTSD2SSrr:
3872 case X86::CVTSS2SDrr:
3873 case X86::Int_CVTSS2SDrr:
3874 case X86::RCPSSr:
3875 case X86::RCPSSr_Int:
3876 case X86::ROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3877 case X86::ROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3878 case X86::ROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3879 case X86::ROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3880 case X86::RSQRTSSr:
3881 case X86::RSQRTSSr_Int:
3882 case X86::SQRTSSr:
3883 case X86::SQRTSSr_Int:
3884 // AVX encoded versions
3885 case X86::VCVTSD2SSrr:
3886 case X86::Int_VCVTSD2SSrr:
3887 case X86::VCVTSS2SDrr:
3888 case X86::Int_VCVTSS2SDrr:
3889 case X86::VRCPSSr:
3890 case X86::VROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3891 case X86::VROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3892 case X86::VROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3893 case X86::VROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3894 case X86::VRSQRTSSr:
3895 case X86::VSQRTSSr:
3896 return true;
3897 }
3898
3899 return false;
3900}
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3901
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]3902/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3903/// instructions we would like before a partial register update.
3904unsigned X86InstrInfo::
3905getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3906 const TargetRegisterInfo *TRI) const {
3907 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3908 return 0;
3909
3910 // If MI is marked as reading Reg, the partial register update is wanted.
3911 const MachineOperand &MO = MI->getOperand(0);
3912 unsigned Reg = MO.getReg();
3913 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3914 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3915 return 0;
3916 } else {
3917 if (MI->readsRegister(Reg, TRI))
3918 return 0;
3919 }
3920
3921 // If any of the preceding 16 instructions are reading Reg, insert a
3922 // dependency breaking instruction. The magic number is based on a few
3923 // Nehalem experiments.
3924 return 16;
3925}
3926
3927void X86InstrInfo::
3928breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3929 const TargetRegisterInfo *TRI) const {
3930 unsigned Reg = MI->getOperand(OpNum).getReg();
3931 if (X86::VR128RegClass.contains(Reg)) {
3932 // These instructions are all floating point domain, so xorps is the best
3933 // choice.
3934 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3935 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3936 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3937 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3938 } else if (X86::VR256RegClass.contains(Reg)) {
3939 // Use vxorps to clear the full ymm register.
3940 // It wants to read and write the xmm sub-register.
3941 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3942 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3943 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3944 .addReg(Reg, RegState::ImplicitDefine);
3945 } else
3946 return;
3947 MI->addRegisterKilled(Reg, TRI, true);
3948}
3949
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3950MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3951 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3952 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3953 int FrameIndex) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3954 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3955 if (NoFusing) return NULL;
3956
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3957 // Unless optimizing for size, don't fold to avoid partial
3958 // register update stalls
Bill Wendling698e84f2012-12-30 10:32:01 +0000[diff] [blame]3959 if (!MF.getFunction()->getAttributes().
3960 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3961 hasPartialRegUpdate(MI->getOpcode()))
3962 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]3963
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3964 const MachineFrameInfo *MFI = MF.getFrameInfo();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3965 unsigned Size = MFI->getObjectSize(FrameIndex);
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3966 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3967 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3968 unsigned NewOpc = 0;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3969 unsigned RCSize = 0;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3970 switch (MI->getOpcode()) {
3971 default: return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3972 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
Dan Gohman887dd1c2010-05-18 21:42:03 +0000[diff] [blame]3973 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
3974 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
3975 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3976 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3977 // Check if it's safe to fold the load. If the size of the object is
3978 // narrower than the load width, then it's not.
3979 if (Size < RCSize)
3980 return NULL;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3981 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]3982 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3983 MI->getOperand(1).ChangeToImmediate(0);
3984 } else if (Ops.size() != 1)
3985 return NULL;
3986
3987 SmallVector<MachineOperand,4> MOs;
3988 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3989 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3990}
3991
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3992MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3993 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3994 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3995 MachineInstr *LoadMI) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3996 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3997 if (NoFusing) return NULL;
3998
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3999 // Unless optimizing for size, don't fold to avoid partial
4000 // register update stalls
Bill Wendling698e84f2012-12-30 10:32:01 +0000[diff] [blame]4001 if (!MF.getFunction()->getAttributes().
4002 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]4003 hasPartialRegUpdate(MI->getOpcode()))
4004 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]4005
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]4006 // Determine the alignment of the load.
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]4007 unsigned Alignment = 0;
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]4008 if (LoadMI->hasOneMemOperand())
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]4009 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4010 else
4011 switch (LoadMI->getOpcode()) {
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]4012 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4013 case X86::AVX_SET0:
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4014 Alignment = 32;
4015 break;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4016 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4017 case X86::V_SETALLONES:
4018 Alignment = 16;
4019 break;
4020 case X86::FsFLD0SD:
4021 Alignment = 8;
4022 break;
4023 case X86::FsFLD0SS:
4024 Alignment = 4;
4025 break;
4026 default:
Eli Friedman87ef3872011-06-10 01:13:01 +0000[diff] [blame]4027 return 0;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4028 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4029 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4030 unsigned NewOpc = 0;
4031 switch (MI->getOpcode()) {
4032 default: return NULL;
4033 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4034 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
4035 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
4036 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4037 }
4038 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4039 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4040 MI->getOperand(1).ChangeToImmediate(0);
4041 } else if (Ops.size() != 1)
4042 return NULL;
4043
Jakob Stoklund Olesen9c473e42010-08-11 23:08:22 +0000[diff] [blame]4044 // Make sure the subregisters match.
4045 // Otherwise we risk changing the size of the load.
4046 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4047 return NULL;
4048
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4049 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4050 switch (LoadMI->getOpcode()) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4051 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4052 case X86::V_SETALLONES:
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]4053 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4054 case X86::AVX_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4055 case X86::FsFLD0SD:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4056 case X86::FsFLD0SS: {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]4057 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4058 // Create a constant-pool entry and operands to load from it.
4059
Dan Gohman772952f2010-03-09 03:01:40 +0000[diff] [blame]4060 // Medium and large mode can't fold loads this way.
4061 if (TM.getCodeModel() != CodeModel::Small &&
4062 TM.getCodeModel() != CodeModel::Kernel)
4063 return NULL;
4064
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4065 // x86-32 PIC requires a PIC base register for constant pools.
4066 unsigned PICBase = 0;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4067 if (TM.getRelocationModel() == Reloc::PIC_) {
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]4068 if (TM.getSubtarget<X86Subtarget>().is64Bit())
4069 PICBase = X86::RIP;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4070 else
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4071 // FIXME: PICBase = getGlobalBaseReg(&MF);
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]4072 // This doesn't work for several reasons.
4073 // 1. GlobalBaseReg may have been spilled.
4074 // 2. It may not be live at MI.
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4075 return NULL;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]4076 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4077
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4078 // Create a constant-pool entry.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4079 MachineConstantPool &MCP = *MF.getConstantPool();
Chris Lattner229907c2011-07-18 04:54:35 +0000[diff] [blame]4080 Type *Ty;
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4081 unsigned Opc = LoadMI->getOpcode();
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4082 if (Opc == X86::FsFLD0SS)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4083 Ty = Type::getFloatTy(MF.getFunction()->getContext());
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]4084 else if (Opc == X86::FsFLD0SD)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4085 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]4086 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
Craig Toppera4c5a472012-01-13 06:12:41 +0000[diff] [blame]4087 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4088 else
4089 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]4090
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]4091 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]4092 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4093 Constant::getNullValue(Ty);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4094 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4095
4096 // Create operands to load from the constant pool entry.
4097 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4098 MOs.push_back(MachineOperand::CreateImm(1));
4099 MOs.push_back(MachineOperand::CreateReg(0, false));
4100 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]4101 MOs.push_back(MachineOperand::CreateReg(0, false));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4102 break;
4103 }
4104 default: {
Manman Ren5b462822012-11-27 18:09:26 +0000[diff] [blame]4105 if ((LoadMI->getOpcode() == X86::MOVSSrm ||
4106 LoadMI->getOpcode() == X86::VMOVSSrm) &&
4107 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4108 > 4)
4109 // These instructions only load 32 bits, we can't fold them if the
4110 // destination register is wider than 32 bits (4 bytes).
4111 return NULL;
4112 if ((LoadMI->getOpcode() == X86::MOVSDrm ||
4113 LoadMI->getOpcode() == X86::VMOVSDrm) &&
4114 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4115 > 8)
4116 // These instructions only load 64 bits, we can't fold them if the
4117 // destination register is wider than 64 bits (8 bytes).
4118 return NULL;
4119
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4120 // Folding a normal load. Just copy the load's address operands.
4121 unsigned NumOps = LoadMI->getDesc().getNumOperands();
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4122 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4123 MOs.push_back(LoadMI->getOperand(i));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]4124 break;
4125 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]4126 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]4127 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4128}
4129
4130
Dan Gohman33332bc2008-10-16 01:49:15 +0000[diff] [blame]4131bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4132 const SmallVectorImpl<unsigned> &Ops) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4133 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4134 if (NoFusing) return 0;
4135
4136 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4137 switch (MI->getOpcode()) {
4138 default: return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4139 case X86::TEST8rr:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4140 case X86::TEST16rr:
4141 case X86::TEST32rr:
4142 case X86::TEST64rr:
4143 return true;
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]4144 case X86::ADD32ri:
4145 // FIXME: AsmPrinter doesn't know how to handle
4146 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4147 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4148 return false;
4149 break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4150 }
4151 }
4152
4153 if (Ops.size() != 1)
4154 return false;
4155
4156 unsigned OpNum = Ops[0];
4157 unsigned Opc = MI->getOpcode();
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]4158 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4159 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4160 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4161
4162 // Folding a memory location into the two-address part of a two-address
4163 // instruction is different than folding it other places. It requires
4164 // replacing the *two* registers with the memory location.
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4165 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4166 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4167 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4168 } else if (OpNum == 0) { // If operand 0
4169 switch (Opc) {
Chris Lattner79c136d2009-07-14 20:19:57 +0000[diff] [blame]4170 case X86::MOV8r0:
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]4171 case X86::MOV16r0:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4172 case X86::MOV32r0:
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4173 case X86::MOV64r0: return true;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4174 default: break;
4175 }
4176 OpcodeTablePtr = &RegOp2MemOpTable0;
4177 } else if (OpNum == 1) {
4178 OpcodeTablePtr = &RegOp2MemOpTable1;
4179 } else if (OpNum == 2) {
4180 OpcodeTablePtr = &RegOp2MemOpTable2;
Craig Topper7573c8f2012-08-31 22:12:16 +0000[diff] [blame]4181 } else if (OpNum == 3) {
4182 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4183 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4184
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]4185 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4186 return true;
Jakob Stoklund Olesen9de596e2012-11-28 02:35:17 +0000[diff] [blame]4187 return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4188}
4189
4190bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4191 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4192 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4193 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4194 MemOp2RegOpTable.find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4195 if (I == MemOp2RegOpTable.end())
4196 return false;
4197 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4198 unsigned Index = I->second.second & TB_INDEX_MASK;
4199 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4200 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4201 if (UnfoldLoad && !FoldedLoad)
4202 return false;
4203 UnfoldLoad &= FoldedLoad;
4204 if (UnfoldStore && !FoldedStore)
4205 return false;
4206 UnfoldStore &= FoldedStore;
4207
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4208 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4209 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4210 if (!MI->hasOneMemOperand() &&
4211 RC == &X86::VR128RegClass &&
4212 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4213 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4214 // conservatively assume the address is unaligned. That's bad for
4215 // performance.
4216 return false;
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4217 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4218 SmallVector<MachineOperand,2> BeforeOps;
4219 SmallVector<MachineOperand,2> AfterOps;
4220 SmallVector<MachineOperand,4> ImpOps;
4221 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4222 MachineOperand &Op = MI->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4223 if (i >= Index && i < Index + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4224 AddrOps.push_back(Op);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4225 else if (Op.isReg() && Op.isImplicit())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4226 ImpOps.push_back(Op);
4227 else if (i < Index)
4228 BeforeOps.push_back(Op);
4229 else if (i > Index)
4230 AfterOps.push_back(Op);
4231 }
4232
4233 // Emit the load instruction.
4234 if (UnfoldLoad) {
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4235 std::pair<MachineInstr::mmo_iterator,
4236 MachineInstr::mmo_iterator> MMOs =
4237 MF.extractLoadMemRefs(MI->memoperands_begin(),
4238 MI->memoperands_end());
4239 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4240 if (UnfoldStore) {
4241 // Address operands cannot be marked isKill.
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4242 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4243 MachineOperand &MO = NewMIs[0]->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4244 if (MO.isReg())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4245 MO.setIsKill(false);
4246 }
4247 }
4248 }
4249
4250 // Emit the data processing instruction.
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4251 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
Jakob Stoklund Olesenb159b5f2012-12-19 21:31:56 +0000[diff] [blame]4252 MachineInstrBuilder MIB(MF, DataMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4253
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4254 if (FoldedStore)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4255 MIB.addReg(Reg, RegState::Define);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4256 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4257 MIB.addOperand(BeforeOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4258 if (FoldedLoad)
4259 MIB.addReg(Reg);
4260 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4261 MIB.addOperand(AfterOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4262 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4263 MachineOperand &MO = ImpOps[i];
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4264 MIB.addReg(MO.getReg(),
4265 getDefRegState(MO.isDef()) |
4266 RegState::Implicit |
4267 getKillRegState(MO.isKill()) |
Evan Cheng0dc101b2009-06-30 08:49:04 +0000[diff] [blame]4268 getDeadRegState(MO.isDead()) |
4269 getUndefRegState(MO.isUndef()));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4270 }
4271 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4272 switch (DataMI->getOpcode()) {
4273 default: break;
4274 case X86::CMP64ri32:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4275 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4276 case X86::CMP32ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4277 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4278 case X86::CMP16ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4279 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4280 case X86::CMP8ri: {
4281 MachineOperand &MO0 = DataMI->getOperand(0);
4282 MachineOperand &MO1 = DataMI->getOperand(1);
4283 if (MO1.getImm() == 0) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4284 unsigned NewOpc;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4285 switch (DataMI->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4286 default: llvm_unreachable("Unreachable!");
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4287 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4288 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4289 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4290 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4291 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4292 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4293 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4294 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4295 DataMI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4296 MO1.ChangeToRegister(MO0.getReg(), false);
4297 }
4298 }
4299 }
4300 NewMIs.push_back(DataMI);
4301
4302 // Emit the store instruction.
4303 if (UnfoldStore) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4304 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4305 std::pair<MachineInstr::mmo_iterator,
4306 MachineInstr::mmo_iterator> MMOs =
4307 MF.extractStoreMemRefs(MI->memoperands_begin(),
4308 MI->memoperands_end());
4309 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4310 }
4311
4312 return true;
4313}
4314
4315bool
4316X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4317 SmallVectorImpl<SDNode*> &NewNodes) const {
Dan Gohman17059682008-07-17 19:10:17 +0000[diff] [blame]4318 if (!N->isMachineOpcode())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4319 return false;
4320
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4321 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4322 MemOp2RegOpTable.find(N->getMachineOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4323 if (I == MemOp2RegOpTable.end())
4324 return false;
4325 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4326 unsigned Index = I->second.second & TB_INDEX_MASK;
4327 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4328 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4329 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4330 MachineFunction &MF = DAG.getMachineFunction();
4331 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4332 unsigned NumDefs = MCID.NumDefs;
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4333 std::vector<SDValue> AddrOps;
4334 std::vector<SDValue> BeforeOps;
4335 std::vector<SDValue> AfterOps;
Dale Johannesen9f3f72f2009-02-06 01:31:28 +0000[diff] [blame]4336 DebugLoc dl = N->getDebugLoc();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4337 unsigned NumOps = N->getNumOperands();
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]4338 for (unsigned i = 0; i != NumOps-1; ++i) {
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4339 SDValue Op = N->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4340 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4341 AddrOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4342 else if (i < Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4343 BeforeOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4344 else if (i > Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4345 AfterOps.push_back(Op);
4346 }
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4347 SDValue Chain = N->getOperand(NumOps-1);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4348 AddrOps.push_back(Chain);
4349
4350 // Emit the load instruction.
4351 SDNode *Load = 0;
4352 if (FoldedLoad) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4353 EVT VT = *RC->vt_begin();
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4354 std::pair<MachineInstr::mmo_iterator,
4355 MachineInstr::mmo_iterator> MMOs =
4356 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4357 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4358 if (!(*MMOs.first) &&
4359 RC == &X86::VR128RegClass &&
4360 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4361 // Do not introduce a slow unaligned load.
4362 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4363 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4364 bool isAligned = (*MMOs.first) &&
4365 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4366 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4367 VT, MVT::Other, AddrOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4368 NewNodes.push_back(Load);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4369
4370 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4371 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4372 }
4373
4374 // Emit the data processing instruction.
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4375 std::vector<EVT> VTs;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4376 const TargetRegisterClass *DstRC = 0;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4377 if (MCID.getNumDefs() > 0) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4378 DstRC = getRegClass(MCID, 0, &RI, MF);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4379 VTs.push_back(*DstRC->vt_begin());
4380 }
4381 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4382 EVT VT = N->getValueType(i);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4383 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4384 VTs.push_back(VT);
4385 }
4386 if (Load)
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4387 BeforeOps.push_back(SDValue(Load, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4388 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4389 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4390 NewNodes.push_back(NewNode);
4391
4392 // Emit the store instruction.
4393 if (FoldedStore) {
4394 AddrOps.pop_back();
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4395 AddrOps.push_back(SDValue(NewNode, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4396 AddrOps.push_back(Chain);
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4397 std::pair<MachineInstr::mmo_iterator,
4398 MachineInstr::mmo_iterator> MMOs =
4399 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4400 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4401 if (!(*MMOs.first) &&
4402 RC == &X86::VR128RegClass &&
4403 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4404 // Do not introduce a slow unaligned store.
4405 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4406 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4407 bool isAligned = (*MMOs.first) &&
4408 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4409 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4410 isAligned, TM),
Michael Liaob53d8962013-04-19 22:22:57 +0000[diff] [blame]4411 dl, MVT::Other, AddrOps);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4412 NewNodes.push_back(Store);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4413
4414 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4415 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4416 }
4417
4418 return true;
4419}
4420
4421unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4422 bool UnfoldLoad, bool UnfoldStore,
4423 unsigned *LoadRegIndex) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4424 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4425 MemOp2RegOpTable.find(Opc);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4426 if (I == MemOp2RegOpTable.end())
4427 return 0;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4428 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4429 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4430 if (UnfoldLoad && !FoldedLoad)
4431 return 0;
4432 if (UnfoldStore && !FoldedStore)
4433 return 0;
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4434 if (LoadRegIndex)
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4435 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4436 return I->second.first;
4437}
4438
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4439bool
4440X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4441 int64_t &Offset1, int64_t &Offset2) const {
4442 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4443 return false;
4444 unsigned Opc1 = Load1->getMachineOpcode();
4445 unsigned Opc2 = Load2->getMachineOpcode();
4446 switch (Opc1) {
4447 default: return false;
4448 case X86::MOV8rm:
4449 case X86::MOV16rm:
4450 case X86::MOV32rm:
4451 case X86::MOV64rm:
4452 case X86::LD_Fp32m:
4453 case X86::LD_Fp64m:
4454 case X86::LD_Fp80m:
4455 case X86::MOVSSrm:
4456 case X86::MOVSDrm:
4457 case X86::MMX_MOVD64rm:
4458 case X86::MMX_MOVQ64rm:
4459 case X86::FsMOVAPSrm:
4460 case X86::FsMOVAPDrm:
4461 case X86::MOVAPSrm:
4462 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4463 case X86::MOVAPDrm:
4464 case X86::MOVDQArm:
4465 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4466 // AVX load instructions
4467 case X86::VMOVSSrm:
4468 case X86::VMOVSDrm:
4469 case X86::FsVMOVAPSrm:
4470 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4471 case X86::VMOVAPSrm:
4472 case X86::VMOVUPSrm:
4473 case X86::VMOVAPDrm:
4474 case X86::VMOVDQArm:
4475 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4476 case X86::VMOVAPSYrm:
4477 case X86::VMOVUPSYrm:
4478 case X86::VMOVAPDYrm:
4479 case X86::VMOVDQAYrm:
4480 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4481 break;
4482 }
4483 switch (Opc2) {
4484 default: return false;
4485 case X86::MOV8rm:
4486 case X86::MOV16rm:
4487 case X86::MOV32rm:
4488 case X86::MOV64rm:
4489 case X86::LD_Fp32m:
4490 case X86::LD_Fp64m:
4491 case X86::LD_Fp80m:
4492 case X86::MOVSSrm:
4493 case X86::MOVSDrm:
4494 case X86::MMX_MOVD64rm:
4495 case X86::MMX_MOVQ64rm:
4496 case X86::FsMOVAPSrm:
4497 case X86::FsMOVAPDrm:
4498 case X86::MOVAPSrm:
4499 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4500 case X86::MOVAPDrm:
4501 case X86::MOVDQArm:
4502 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4503 // AVX load instructions
4504 case X86::VMOVSSrm:
4505 case X86::VMOVSDrm:
4506 case X86::FsVMOVAPSrm:
4507 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4508 case X86::VMOVAPSrm:
4509 case X86::VMOVUPSrm:
4510 case X86::VMOVAPDrm:
4511 case X86::VMOVDQArm:
4512 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4513 case X86::VMOVAPSYrm:
4514 case X86::VMOVUPSYrm:
4515 case X86::VMOVAPDYrm:
4516 case X86::VMOVDQAYrm:
4517 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4518 break;
4519 }
4520
4521 // Check if chain operands and base addresses match.
4522 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4523 Load1->getOperand(5) != Load2->getOperand(5))
4524 return false;
4525 // Segment operands should match as well.
4526 if (Load1->getOperand(4) != Load2->getOperand(4))
4527 return false;
4528 // Scale should be 1, Index should be Reg0.
4529 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4530 Load1->getOperand(2) == Load2->getOperand(2)) {
4531 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4532 return false;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4533
4534 // Now let's examine the displacements.
4535 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4536 isa<ConstantSDNode>(Load2->getOperand(3))) {
4537 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4538 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4539 return true;
4540 }
4541 }
4542 return false;
4543}
4544
4545bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4546 int64_t Offset1, int64_t Offset2,
4547 unsigned NumLoads) const {
4548 assert(Offset2 > Offset1);
4549 if ((Offset2 - Offset1) / 8 > 64)
4550 return false;
4551
4552 unsigned Opc1 = Load1->getMachineOpcode();
4553 unsigned Opc2 = Load2->getMachineOpcode();
4554 if (Opc1 != Opc2)
4555 return false; // FIXME: overly conservative?
4556
4557 switch (Opc1) {
4558 default: break;
4559 case X86::LD_Fp32m:
4560 case X86::LD_Fp64m:
4561 case X86::LD_Fp80m:
4562 case X86::MMX_MOVD64rm:
4563 case X86::MMX_MOVQ64rm:
4564 return false;
4565 }
4566
4567 EVT VT = Load1->getValueType(0);
4568 switch (VT.getSimpleVT().SimpleTy) {
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4569 default:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4570 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4571 // have 16 of them to play with.
4572 if (TM.getSubtargetImpl()->is64Bit()) {
4573 if (NumLoads >= 3)
4574 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4575 } else if (NumLoads) {
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4576 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4577 }
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4578 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4579 case MVT::i8:
4580 case MVT::i16:
4581 case MVT::i32:
4582 case MVT::i64:
Evan Cheng16cf9342010-01-22 23:49:11 +0000[diff] [blame]4583 case MVT::f32:
4584 case MVT::f64:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4585 if (NumLoads)
4586 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4587 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4588 }
4589
4590 return true;
4591}
4592
4593
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4594bool X86InstrInfo::
Owen Anderson4f6bf042008-08-14 22:49:33 +0000[diff] [blame]4595ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4596 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4597 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]4598 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4599 return true;
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4600 Cond[0].setImm(GetOppositeBranchCondition(CC));
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4601 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4602}
4603
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4604bool X86InstrInfo::
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4605isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4606 // FIXME: Return false for x87 stack register classes for now. We can't
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4607 // allow any loads of these registers before FpGet_ST0_80.
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4608 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4609 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4610}
4611
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4612/// getGlobalBaseReg - Return a virtual register initialized with the
4613/// the global base register value. Output instructions required to
4614/// initialize the register in the function entry block, if necessary.
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4615///
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4616/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4617///
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4618unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4619 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4620 "X86-64 PIC uses RIP relative addressing");
4621
4622 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4623 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4624 if (GlobalBaseReg != 0)
4625 return GlobalBaseReg;
4626
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4627 // Create the register. The code to initialize it is inserted
4628 // later, by the CGBR pass (below).
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4629 MachineRegisterInfo &RegInfo = MF->getRegInfo();
Jakob Stoklund Olesen38dcd592012-05-20 18:43:00 +0000[diff] [blame]4630 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4631 X86FI->setGlobalBaseReg(GlobalBaseReg);
4632 return GlobalBaseReg;
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4633}
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4634
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4635// These are the replaceable SSE instructions. Some of these have Int variants
4636// that we don't include here. We don't want to replace instructions selected
4637// by intrinsics.
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4638static const uint16_t ReplaceableInstrs[][3] = {
Bruno Cardoso Lopes1401e042010-08-12 02:08:52 +0000[diff] [blame]4639 //PackedSingle PackedDouble PackedInt
Jakob Stoklund Olesendbff4e82010-03-30 22:46:53 +0000[diff] [blame]4640 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4641 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4642 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4643 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4644 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4645 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4646 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4647 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4648 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4649 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4650 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4651 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4652 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4653 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4654 // AVX 128-bit support
4655 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4656 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4657 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4658 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4659 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4660 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4661 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4662 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4663 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4664 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4665 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4666 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4667 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4668 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4669 // AVX 256-bit support
4670 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4671 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4672 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4673 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4674 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4675 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4676};
4677
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4678static const uint16_t ReplaceableInstrsAVX2[][3] = {
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4679 //PackedSingle PackedDouble PackedInt
Craig Topperf87a2be2011-11-09 09:37:21 +0000[diff] [blame]4680 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4681 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4682 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4683 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4684 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4685 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4686 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
Craig Topper12b72de2011-11-29 05:37:58 +0000[diff] [blame]4687 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4688 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4689 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4690 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4691 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4692 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4693 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4694};
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4695
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4696// FIXME: Some shuffle and unpack instructions have equivalents in different
4697// domains, but they require a bit more work than just switching opcodes.
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4698
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4699static const uint16_t *lookup(unsigned opcode, unsigned domain) {
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4700 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4701 if (ReplaceableInstrs[i][domain-1] == opcode)
4702 return ReplaceableInstrs[i];
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4703 return 0;
4704}
4705
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4706static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4707 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4708 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4709 return ReplaceableInstrsAVX2[i];
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4710 return 0;
4711}
4712
4713std::pair<uint16_t, uint16_t>
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]4714X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4715 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4716 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4717 uint16_t validDomains = 0;
4718 if (domain && lookup(MI->getOpcode(), domain))
4719 validDomains = 0xe;
4720 else if (domain && lookupAVX2(MI->getOpcode(), domain))
4721 validDomains = hasAVX2 ? 0xe : 0x6;
4722 return std::make_pair(domain, validDomains);
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4723}
4724
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]4725void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4726 assert(Domain>0 && Domain<4 && "Invalid execution domain");
4727 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4728 assert(dom && "Not an SSE instruction");
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4729 const uint16_t *table = lookup(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]4730 if (!table) { // try the other table
4731 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
4732 "256-bit vector operations only available in AVX2");
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4733 table = lookupAVX2(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]4734 }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4735 assert(table && "Cannot change domain");
4736 MI->setDesc(get(table[Domain-1]));
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4737}
Chris Lattner6a5e7062010-04-26 23:37:21 +0000[diff] [blame]4738
4739/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
4740void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
4741 NopInst.setOpcode(X86::NOOP);
4742}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4743
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]4744bool X86InstrInfo::isHighLatencyDef(int opc) const {
4745 switch (opc) {
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4746 default: return false;
4747 case X86::DIVSDrm:
4748 case X86::DIVSDrm_Int:
4749 case X86::DIVSDrr:
4750 case X86::DIVSDrr_Int:
4751 case X86::DIVSSrm:
4752 case X86::DIVSSrm_Int:
4753 case X86::DIVSSrr:
4754 case X86::DIVSSrr_Int:
4755 case X86::SQRTPDm:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4756 case X86::SQRTPDr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4757 case X86::SQRTPSm:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4758 case X86::SQRTPSr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4759 case X86::SQRTSDm:
4760 case X86::SQRTSDm_Int:
4761 case X86::SQRTSDr:
4762 case X86::SQRTSDr_Int:
4763 case X86::SQRTSSm:
4764 case X86::SQRTSSm_Int:
4765 case X86::SQRTSSr:
4766 case X86::SQRTSSr_Int:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4767 // AVX instructions with high latency
4768 case X86::VDIVSDrm:
4769 case X86::VDIVSDrm_Int:
4770 case X86::VDIVSDrr:
4771 case X86::VDIVSDrr_Int:
4772 case X86::VDIVSSrm:
4773 case X86::VDIVSSrm_Int:
4774 case X86::VDIVSSrr:
4775 case X86::VDIVSSrr_Int:
4776 case X86::VSQRTPDm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4777 case X86::VSQRTPDr:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4778 case X86::VSQRTPSm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4779 case X86::VSQRTPSr:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4780 case X86::VSQRTSDm:
4781 case X86::VSQRTSDm_Int:
4782 case X86::VSQRTSDr:
4783 case X86::VSQRTSSm:
4784 case X86::VSQRTSSm_Int:
4785 case X86::VSQRTSSr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4786 return true;
4787 }
4788}
4789
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]4790bool X86InstrInfo::
4791hasHighOperandLatency(const InstrItineraryData *ItinData,
4792 const MachineRegisterInfo *MRI,
4793 const MachineInstr *DefMI, unsigned DefIdx,
4794 const MachineInstr *UseMI, unsigned UseIdx) const {
4795 return isHighLatencyDef(DefMI->getOpcode());
4796}
4797
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4798namespace {
4799 /// CGBR - Create Global Base Reg pass. This initializes the PIC
4800 /// global base register for x86-32.
4801 struct CGBR : public MachineFunctionPass {
4802 static char ID;
Owen Andersona7aed182010-08-06 18:33:48 +0000[diff] [blame]4803 CGBR() : MachineFunctionPass(ID) {}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4804
4805 virtual bool runOnMachineFunction(MachineFunction &MF) {
4806 const X86TargetMachine *TM =
4807 static_cast<const X86TargetMachine *>(&MF.getTarget());
4808
4809 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
4810 "X86-64 PIC uses RIP relative addressing");
4811
4812 // Only emit a global base reg in PIC mode.
4813 if (TM->getRelocationModel() != Reloc::PIC_)
4814 return false;
4815
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]4816 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
4817 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4818
4819 // If we didn't need a GlobalBaseReg, don't insert code.
4820 if (GlobalBaseReg == 0)
4821 return false;
4822
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4823 // Insert the set of GlobalBaseReg into the first MBB of the function
4824 MachineBasicBlock &FirstMBB = MF.front();
4825 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
4826 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
4827 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4828 const X86InstrInfo *TII = TM->getInstrInfo();
4829
4830 unsigned PC;
4831 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]4832 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4833 else
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]4834 PC = GlobalBaseReg;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4835
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4836 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4837 // only used in JIT code emission as displacement to pc.
4838 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4839
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4840 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4841 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4842 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4843 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4844 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4845 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4846 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4847 }
4848
4849 return true;
4850 }
4851
4852 virtual const char *getPassName() const {
4853 return "X86 PIC Global Base Reg Initialization";
4854 }
4855
4856 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4857 AU.setPreservesCFG();
4858 MachineFunctionPass::getAnalysisUsage(AU);
4859 }
4860 };
4861}
4862
4863char CGBR::ID = 0;
4864FunctionPass*
4865llvm::createGlobalBaseRegPass() { return new CGBR(); }
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]4866
4867namespace {
4868 struct LDTLSCleanup : public MachineFunctionPass {
4869 static char ID;
4870 LDTLSCleanup() : MachineFunctionPass(ID) {}
4871
4872 virtual bool runOnMachineFunction(MachineFunction &MF) {
4873 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4874 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4875 // No point folding accesses if there isn't at least two.
4876 return false;
4877 }
4878
4879 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4880 return VisitNode(DT->getRootNode(), 0);
4881 }
4882
4883 // Visit the dominator subtree rooted at Node in pre-order.
4884 // If TLSBaseAddrReg is non-null, then use that to replace any
4885 // TLS_base_addr instructions. Otherwise, create the register
4886 // when the first such instruction is seen, and then use it
4887 // as we encounter more instructions.
4888 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4889 MachineBasicBlock *BB = Node->getBlock();
4890 bool Changed = false;
4891
4892 // Traverse the current block.
4893 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4894 ++I) {
4895 switch (I->getOpcode()) {
4896 case X86::TLS_base_addr32:
4897 case X86::TLS_base_addr64:
4898 if (TLSBaseAddrReg)
4899 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4900 else
4901 I = SetRegister(I, &TLSBaseAddrReg);
4902 Changed = true;
4903 break;
4904 default:
4905 break;
4906 }
4907 }
4908
4909 // Visit the children of this block in the dominator tree.
4910 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4911 I != E; ++I) {
4912 Changed |= VisitNode(*I, TLSBaseAddrReg);
4913 }
4914
4915 return Changed;
4916 }
4917
4918 // Replace the TLS_base_addr instruction I with a copy from
4919 // TLSBaseAddrReg, returning the new instruction.
4920 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4921 unsigned TLSBaseAddrReg) {
4922 MachineFunction *MF = I->getParent()->getParent();
4923 const X86TargetMachine *TM =
4924 static_cast<const X86TargetMachine *>(&MF->getTarget());
4925 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4926 const X86InstrInfo *TII = TM->getInstrInfo();
4927
4928 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4929 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4930 TII->get(TargetOpcode::COPY),
4931 is64Bit ? X86::RAX : X86::EAX)
4932 .addReg(TLSBaseAddrReg);
4933
4934 // Erase the TLS_base_addr instruction.
4935 I->eraseFromParent();
4936
4937 return Copy;
4938 }
4939
4940 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4941 // inserting a copy instruction after I. Returns the new instruction.
4942 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4943 MachineFunction *MF = I->getParent()->getParent();
4944 const X86TargetMachine *TM =
4945 static_cast<const X86TargetMachine *>(&MF->getTarget());
4946 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4947 const X86InstrInfo *TII = TM->getInstrInfo();
4948
4949 // Create a virtual register for the TLS base address.
4950 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4951 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4952 ? &X86::GR64RegClass
4953 : &X86::GR32RegClass);
4954
4955 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4956 MachineInstr *Next = I->getNextNode();
4957 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4958 TII->get(TargetOpcode::COPY),
4959 *TLSBaseAddrReg)
4960 .addReg(is64Bit ? X86::RAX : X86::EAX);
4961
4962 return Copy;
4963 }
4964
4965 virtual const char *getPassName() const {
4966 return "Local Dynamic TLS Access Clean-up";
4967 }
4968
4969 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4970 AU.setPreservesCFG();
4971 AU.addRequired<MachineDominatorTree>();
4972 MachineFunctionPass::getAnalysisUsage(AU);
4973 }
4974 };
4975}
4976
4977char LDTLSCleanup::ID = 0;
4978FunctionPass*
4979llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }