Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / 2de86af22d3e581ee1e82b2f14de2cc8b4ddaacb / . / llvm / lib / Target / X86 / X86InstrInfo.cpp
blob: 2fb2ed2104c74ea3b828cd479d024a74deae6980 [file] [log] [blame]
Jia Liub22310f2012-02-18 12:03:15 +0000[diff] [blame]1//===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
Misha Brukmanc88330a2005-04-21 23:38:14 +0000[diff] [blame]2//
John Criswell482202a2003-10-20 19:43:21 +0000[diff] [blame]3// The LLVM Compiler Infrastructure
4//
Chris Lattnerf3ebc3f2007-12-29 20:36:04 +0000[diff] [blame]5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
Misha Brukmanc88330a2005-04-21 23:38:14 +0000[diff] [blame]7//
John Criswell482202a2003-10-20 19:43:21 +0000[diff] [blame]8//===----------------------------------------------------------------------===//
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]9//
Chris Lattnerb4d58d72003-01-14 22:00:31 +0000[diff] [blame]10// This file contains the X86 implementation of the TargetInstrInfo class.
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]11//
12//===----------------------------------------------------------------------===//
13
Chris Lattner27d24792002-10-29 21:05:24 +0000[diff] [blame]14#include "X86InstrInfo.h"
Chris Lattner0d808742002-12-03 05:42:53 +0000[diff] [blame]15#include "X86.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]16#include "X86InstrBuilder.h"
Owen Anderson6bb0c522008-01-04 23:57:37 +0000[diff] [blame]17#include "X86MachineFunctionInfo.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]18#include "X86Subtarget.h"
19#include "X86TargetMachine.h"
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]20#include "llvm/DerivedTypes.h"
Owen Anderson53a52212009-07-13 04:09:18 +0000[diff] [blame]21#include "llvm/LLVMContext.h"
Owen Andersone2f23a32007-09-07 04:06:50 +0000[diff] [blame]22#include "llvm/ADT/STLExtras.h"
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]23#include "llvm/CodeGen/MachineConstantPool.h"
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]24#include "llvm/CodeGen/MachineDominators.h"
Owen Anderson6bb0c522008-01-04 23:57:37 +0000[diff] [blame]25#include "llvm/CodeGen/MachineFrameInfo.h"
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]26#include "llvm/CodeGen/MachineInstrBuilder.h"
Chris Lattnera10fff52007-12-31 04:13:23 +0000[diff] [blame]27#include "llvm/CodeGen/MachineRegisterInfo.h"
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]28#include "llvm/CodeGen/LiveVariables.h"
Craig Topperb25fda92012-03-17 18:46:09 +0000[diff] [blame]29#include "llvm/MC/MCAsmInfo.h"
Chris Lattner6a5e7062010-04-26 23:37:21 +0000[diff] [blame]30#include "llvm/MC/MCInst.h"
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]31#include "llvm/Support/CommandLine.h"
David Greened589daf2010-01-05 01:29:29 +0000[diff] [blame]32#include "llvm/Support/Debug.h"
Torok Edwin6dd27302009-07-08 18:01:40 +0000[diff] [blame]33#include "llvm/Support/ErrorHandling.h"
34#include "llvm/Support/raw_ostream.h"
Evan Chenge95f3912007-09-25 01:57:46 +0000[diff] [blame]35#include "llvm/Target/TargetOptions.h"
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]36#include <limits>
37
Evan Cheng703a0fb2011-07-01 17:57:27 +0000[diff] [blame]38#define GET_INSTRINFO_CTOR
Evan Cheng1e210d02011-06-28 20:07:07 +0000[diff] [blame]39#include "X86GenInstrInfo.inc"
40
Brian Gaeke960707c2003-11-11 22:41:34 +0000[diff] [blame]41using namespace llvm;
42
Chris Lattnera6f074f2009-08-23 03:41:05 +0000[diff] [blame]43static cl::opt<bool>
44NoFusing("disable-spill-fusing",
45 cl::desc("Disable fusing of spill code into instructions"));
46static cl::opt<bool>
47PrintFailedFusing("print-failed-fuse-candidates",
48 cl::desc("Print instructions that the allocator wants to"
49 " fuse, but the X86 backend currently can't"),
50 cl::Hidden);
51static cl::opt<bool>
52ReMatPICStubLoad("remat-pic-stub-load",
53 cl::desc("Re-materialize load from stub in PIC mode"),
54 cl::init(false), cl::Hidden);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]55
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]56enum {
57 // Select which memory operand is being unfolded.
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]58 // (stored in bits 0 - 3)
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]59 TB_INDEX_0 = 0,
60 TB_INDEX_1 = 1,
61 TB_INDEX_2 = 2,
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]62 TB_INDEX_3 = 3,
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]63 TB_INDEX_MASK = 0xf,
64
65 // Do not insert the reverse map (MemOp -> RegOp) into the table.
66 // This may be needed because there is a many -> one mapping.
67 TB_NO_REVERSE = 1 << 4,
68
69 // Do not insert the forward map (RegOp -> MemOp) into the table.
70 // This is needed for Native Client, which prohibits branch
71 // instructions from using a memory operand.
72 TB_NO_FORWARD = 1 << 5,
73
74 TB_FOLDED_LOAD = 1 << 6,
75 TB_FOLDED_STORE = 1 << 7,
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]76
77 // Minimum alignment required for load/store.
78 // Used for RegOp->MemOp conversion.
79 // (stored in bits 8 - 15)
80 TB_ALIGN_SHIFT = 8,
81 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
82 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
83 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]84 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]85};
86
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]87struct X86OpTblEntry {
88 uint16_t RegOp;
89 uint16_t MemOp;
Craig Topper1cac50b2012-06-23 08:01:18 +0000[diff] [blame]90 uint16_t Flags;
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]91};
92
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]93X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
Evan Cheng703a0fb2011-07-01 17:57:27 +0000[diff] [blame]94 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
95 ? X86::ADJCALLSTACKDOWN64
96 : X86::ADJCALLSTACKDOWN32),
97 (tm.getSubtarget<X86Subtarget>().is64Bit()
98 ? X86::ADJCALLSTACKUP64
99 : X86::ADJCALLSTACKUP32)),
Evan Cheng11b0a5d2006-09-08 06:48:29 +0000[diff] [blame]100 TM(tm), RI(tm, *this) {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]101
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]102 static const X86OpTblEntry OpTbl2Addr[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]103 { X86::ADC32ri, X86::ADC32mi, 0 },
104 { X86::ADC32ri8, X86::ADC32mi8, 0 },
105 { X86::ADC32rr, X86::ADC32mr, 0 },
106 { X86::ADC64ri32, X86::ADC64mi32, 0 },
107 { X86::ADC64ri8, X86::ADC64mi8, 0 },
108 { X86::ADC64rr, X86::ADC64mr, 0 },
109 { X86::ADD16ri, X86::ADD16mi, 0 },
110 { X86::ADD16ri8, X86::ADD16mi8, 0 },
111 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
112 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
113 { X86::ADD16rr, X86::ADD16mr, 0 },
114 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
115 { X86::ADD32ri, X86::ADD32mi, 0 },
116 { X86::ADD32ri8, X86::ADD32mi8, 0 },
117 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
118 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
119 { X86::ADD32rr, X86::ADD32mr, 0 },
120 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
121 { X86::ADD64ri32, X86::ADD64mi32, 0 },
122 { X86::ADD64ri8, X86::ADD64mi8, 0 },
123 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
124 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
125 { X86::ADD64rr, X86::ADD64mr, 0 },
126 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
127 { X86::ADD8ri, X86::ADD8mi, 0 },
128 { X86::ADD8rr, X86::ADD8mr, 0 },
129 { X86::AND16ri, X86::AND16mi, 0 },
130 { X86::AND16ri8, X86::AND16mi8, 0 },
131 { X86::AND16rr, X86::AND16mr, 0 },
132 { X86::AND32ri, X86::AND32mi, 0 },
133 { X86::AND32ri8, X86::AND32mi8, 0 },
134 { X86::AND32rr, X86::AND32mr, 0 },
135 { X86::AND64ri32, X86::AND64mi32, 0 },
136 { X86::AND64ri8, X86::AND64mi8, 0 },
137 { X86::AND64rr, X86::AND64mr, 0 },
138 { X86::AND8ri, X86::AND8mi, 0 },
139 { X86::AND8rr, X86::AND8mr, 0 },
140 { X86::DEC16r, X86::DEC16m, 0 },
141 { X86::DEC32r, X86::DEC32m, 0 },
142 { X86::DEC64_16r, X86::DEC64_16m, 0 },
143 { X86::DEC64_32r, X86::DEC64_32m, 0 },
144 { X86::DEC64r, X86::DEC64m, 0 },
145 { X86::DEC8r, X86::DEC8m, 0 },
146 { X86::INC16r, X86::INC16m, 0 },
147 { X86::INC32r, X86::INC32m, 0 },
148 { X86::INC64_16r, X86::INC64_16m, 0 },
149 { X86::INC64_32r, X86::INC64_32m, 0 },
150 { X86::INC64r, X86::INC64m, 0 },
151 { X86::INC8r, X86::INC8m, 0 },
152 { X86::NEG16r, X86::NEG16m, 0 },
153 { X86::NEG32r, X86::NEG32m, 0 },
154 { X86::NEG64r, X86::NEG64m, 0 },
155 { X86::NEG8r, X86::NEG8m, 0 },
156 { X86::NOT16r, X86::NOT16m, 0 },
157 { X86::NOT32r, X86::NOT32m, 0 },
158 { X86::NOT64r, X86::NOT64m, 0 },
159 { X86::NOT8r, X86::NOT8m, 0 },
160 { X86::OR16ri, X86::OR16mi, 0 },
161 { X86::OR16ri8, X86::OR16mi8, 0 },
162 { X86::OR16rr, X86::OR16mr, 0 },
163 { X86::OR32ri, X86::OR32mi, 0 },
164 { X86::OR32ri8, X86::OR32mi8, 0 },
165 { X86::OR32rr, X86::OR32mr, 0 },
166 { X86::OR64ri32, X86::OR64mi32, 0 },
167 { X86::OR64ri8, X86::OR64mi8, 0 },
168 { X86::OR64rr, X86::OR64mr, 0 },
169 { X86::OR8ri, X86::OR8mi, 0 },
170 { X86::OR8rr, X86::OR8mr, 0 },
171 { X86::ROL16r1, X86::ROL16m1, 0 },
172 { X86::ROL16rCL, X86::ROL16mCL, 0 },
173 { X86::ROL16ri, X86::ROL16mi, 0 },
174 { X86::ROL32r1, X86::ROL32m1, 0 },
175 { X86::ROL32rCL, X86::ROL32mCL, 0 },
176 { X86::ROL32ri, X86::ROL32mi, 0 },
177 { X86::ROL64r1, X86::ROL64m1, 0 },
178 { X86::ROL64rCL, X86::ROL64mCL, 0 },
179 { X86::ROL64ri, X86::ROL64mi, 0 },
180 { X86::ROL8r1, X86::ROL8m1, 0 },
181 { X86::ROL8rCL, X86::ROL8mCL, 0 },
182 { X86::ROL8ri, X86::ROL8mi, 0 },
183 { X86::ROR16r1, X86::ROR16m1, 0 },
184 { X86::ROR16rCL, X86::ROR16mCL, 0 },
185 { X86::ROR16ri, X86::ROR16mi, 0 },
186 { X86::ROR32r1, X86::ROR32m1, 0 },
187 { X86::ROR32rCL, X86::ROR32mCL, 0 },
188 { X86::ROR32ri, X86::ROR32mi, 0 },
189 { X86::ROR64r1, X86::ROR64m1, 0 },
190 { X86::ROR64rCL, X86::ROR64mCL, 0 },
191 { X86::ROR64ri, X86::ROR64mi, 0 },
192 { X86::ROR8r1, X86::ROR8m1, 0 },
193 { X86::ROR8rCL, X86::ROR8mCL, 0 },
194 { X86::ROR8ri, X86::ROR8mi, 0 },
195 { X86::SAR16r1, X86::SAR16m1, 0 },
196 { X86::SAR16rCL, X86::SAR16mCL, 0 },
197 { X86::SAR16ri, X86::SAR16mi, 0 },
198 { X86::SAR32r1, X86::SAR32m1, 0 },
199 { X86::SAR32rCL, X86::SAR32mCL, 0 },
200 { X86::SAR32ri, X86::SAR32mi, 0 },
201 { X86::SAR64r1, X86::SAR64m1, 0 },
202 { X86::SAR64rCL, X86::SAR64mCL, 0 },
203 { X86::SAR64ri, X86::SAR64mi, 0 },
204 { X86::SAR8r1, X86::SAR8m1, 0 },
205 { X86::SAR8rCL, X86::SAR8mCL, 0 },
206 { X86::SAR8ri, X86::SAR8mi, 0 },
207 { X86::SBB32ri, X86::SBB32mi, 0 },
208 { X86::SBB32ri8, X86::SBB32mi8, 0 },
209 { X86::SBB32rr, X86::SBB32mr, 0 },
210 { X86::SBB64ri32, X86::SBB64mi32, 0 },
211 { X86::SBB64ri8, X86::SBB64mi8, 0 },
212 { X86::SBB64rr, X86::SBB64mr, 0 },
213 { X86::SHL16rCL, X86::SHL16mCL, 0 },
214 { X86::SHL16ri, X86::SHL16mi, 0 },
215 { X86::SHL32rCL, X86::SHL32mCL, 0 },
216 { X86::SHL32ri, X86::SHL32mi, 0 },
217 { X86::SHL64rCL, X86::SHL64mCL, 0 },
218 { X86::SHL64ri, X86::SHL64mi, 0 },
219 { X86::SHL8rCL, X86::SHL8mCL, 0 },
220 { X86::SHL8ri, X86::SHL8mi, 0 },
221 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
222 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
223 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
224 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
225 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
226 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
227 { X86::SHR16r1, X86::SHR16m1, 0 },
228 { X86::SHR16rCL, X86::SHR16mCL, 0 },
229 { X86::SHR16ri, X86::SHR16mi, 0 },
230 { X86::SHR32r1, X86::SHR32m1, 0 },
231 { X86::SHR32rCL, X86::SHR32mCL, 0 },
232 { X86::SHR32ri, X86::SHR32mi, 0 },
233 { X86::SHR64r1, X86::SHR64m1, 0 },
234 { X86::SHR64rCL, X86::SHR64mCL, 0 },
235 { X86::SHR64ri, X86::SHR64mi, 0 },
236 { X86::SHR8r1, X86::SHR8m1, 0 },
237 { X86::SHR8rCL, X86::SHR8mCL, 0 },
238 { X86::SHR8ri, X86::SHR8mi, 0 },
239 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
240 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
241 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
242 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
243 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
244 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
245 { X86::SUB16ri, X86::SUB16mi, 0 },
246 { X86::SUB16ri8, X86::SUB16mi8, 0 },
247 { X86::SUB16rr, X86::SUB16mr, 0 },
248 { X86::SUB32ri, X86::SUB32mi, 0 },
249 { X86::SUB32ri8, X86::SUB32mi8, 0 },
250 { X86::SUB32rr, X86::SUB32mr, 0 },
251 { X86::SUB64ri32, X86::SUB64mi32, 0 },
252 { X86::SUB64ri8, X86::SUB64mi8, 0 },
253 { X86::SUB64rr, X86::SUB64mr, 0 },
254 { X86::SUB8ri, X86::SUB8mi, 0 },
255 { X86::SUB8rr, X86::SUB8mr, 0 },
256 { X86::XOR16ri, X86::XOR16mi, 0 },
257 { X86::XOR16ri8, X86::XOR16mi8, 0 },
258 { X86::XOR16rr, X86::XOR16mr, 0 },
259 { X86::XOR32ri, X86::XOR32mi, 0 },
260 { X86::XOR32ri8, X86::XOR32mi8, 0 },
261 { X86::XOR32rr, X86::XOR32mr, 0 },
262 { X86::XOR64ri32, X86::XOR64mi32, 0 },
263 { X86::XOR64ri8, X86::XOR64mi8, 0 },
264 { X86::XOR64rr, X86::XOR64mr, 0 },
265 { X86::XOR8ri, X86::XOR8mi, 0 },
266 { X86::XOR8rr, X86::XOR8mr, 0 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]267 };
268
269 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]270 unsigned RegOp = OpTbl2Addr[i].RegOp;
271 unsigned MemOp = OpTbl2Addr[i].MemOp;
272 unsigned Flags = OpTbl2Addr[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]273 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
274 RegOp, MemOp,
275 // Index 0, folded load and store, no alignment requirement.
276 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]277 }
278
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]279 static const X86OpTblEntry OpTbl0[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]280 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
281 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
282 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
283 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
284 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]285 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
286 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
287 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
288 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
289 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
290 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
291 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
292 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
293 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
294 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
295 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
296 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
297 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
298 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
299 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
300 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
301 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
302 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]303 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
304 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
305 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
306 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
307 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
308 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
309 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
310 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
311 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
312 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
313 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
314 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
315 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
316 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
317 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
318 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
319 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
320 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
321 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
322 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
323 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
324 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]325 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
326 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
327 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
328 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
329 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
330 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]331 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
332 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
333 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
334 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
335 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
336 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
337 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
338 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
339 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
340 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
341 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
342 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
343 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
344 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
345 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
346 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
347 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
348 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
349 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
350 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
351 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
352 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
353 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
354 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
355 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]356 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
357 // AVX 128-bit versions of foldable instructions
358 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
359 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
360 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]361 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]362 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
363 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
366 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
367 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
368 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
369 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
370 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
371 // AVX 256-bit foldable instructions
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]372 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]373 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
374 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
377 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]378 };
379
380 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]381 unsigned RegOp = OpTbl0[i].RegOp;
382 unsigned MemOp = OpTbl0[i].MemOp;
383 unsigned Flags = OpTbl0[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]384 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
385 RegOp, MemOp, TB_INDEX_0 | Flags);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]386 }
387
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]388 static const X86OpTblEntry OpTbl1[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]389 { X86::CMP16rr, X86::CMP16rm, 0 },
390 { X86::CMP32rr, X86::CMP32rm, 0 },
391 { X86::CMP64rr, X86::CMP64rm, 0 },
392 { X86::CMP8rr, X86::CMP8rm, 0 },
393 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
394 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
395 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
396 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
397 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
398 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
399 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
400 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
401 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
402 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
403 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
404 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]405 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
406 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
407 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
408 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
409 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
410 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
411 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
412 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]413 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
414 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]415 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
416 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]417 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
418 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
419 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
420 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
421 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
422 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
423 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
424 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]425 { X86::MOV16rr, X86::MOV16rm, 0 },
426 { X86::MOV32rr, X86::MOV32rm, 0 },
427 { X86::MOV64rr, X86::MOV64rm, 0 },
428 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
429 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
430 { X86::MOV8rr, X86::MOV8rm, 0 },
431 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
432 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]433 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
434 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
435 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
436 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]437 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
438 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
439 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
440 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
441 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
442 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
443 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
444 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
445 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
446 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]447 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
448 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
449 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
450 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
451 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
452 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
453 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
454 { X86::MOVZX64rr16, X86::MOVZX64rm16, 0 },
455 { X86::MOVZX64rr32, X86::MOVZX64rm32, 0 },
456 { X86::MOVZX64rr8, X86::MOVZX64rm8, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]457 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
458 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
459 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]460 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
461 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
462 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
463 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
464 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
465 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
466 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
467 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
468 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
469 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
470 { X86::SQRTPDr_Int, X86::SQRTPDm_Int, TB_ALIGN_16 },
471 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
472 { X86::SQRTPSr_Int, X86::SQRTPSm_Int, TB_ALIGN_16 },
473 { X86::SQRTSDr, X86::SQRTSDm, 0 },
474 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
475 { X86::SQRTSSr, X86::SQRTSSm, 0 },
476 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
477 { X86::TEST16rr, X86::TEST16rm, 0 },
478 { X86::TEST32rr, X86::TEST32rm, 0 },
479 { X86::TEST64rr, X86::TEST64rm, 0 },
480 { X86::TEST8rr, X86::TEST8rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]481 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]482 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
483 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]484 // AVX 128-bit versions of foldable instructions
485 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
486 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]487 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
488 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]489 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
490 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
Pete Cooper8bbce762012-06-14 22:12:58 +0000[diff] [blame]491 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
Craig Topper11913052012-06-15 07:02:58 +0000[diff] [blame]492 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
493 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
494 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
495 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
496 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
497 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
498 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
499 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
500 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]501 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
502 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
503 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
504 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
505 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
506 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
507 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
508 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
509 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
510 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
511 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
512 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
513 { X86::VMOVUPDrr, X86::VMOVUPDrm, TB_ALIGN_16 },
514 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
515 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
516 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
517 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]518 { X86::VPABSBrr128, X86::VPABSBrm128, TB_ALIGN_16 },
519 { X86::VPABSDrr128, X86::VPABSDrm128, TB_ALIGN_16 },
520 { X86::VPABSWrr128, X86::VPABSWrm128, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]521 { X86::VPERMILPDri, X86::VPERMILPDmi, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]522 { X86::VPERMILPSri, X86::VPERMILPSmi, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]523 { X86::VPSHUFDri, X86::VPSHUFDmi, TB_ALIGN_16 },
524 { X86::VPSHUFHWri, X86::VPSHUFHWmi, TB_ALIGN_16 },
525 { X86::VPSHUFLWri, X86::VPSHUFLWmi, TB_ALIGN_16 },
526 { X86::VRCPPSr, X86::VRCPPSm, TB_ALIGN_16 },
527 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, TB_ALIGN_16 },
528 { X86::VRSQRTPSr, X86::VRSQRTPSm, TB_ALIGN_16 },
529 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, TB_ALIGN_16 },
530 { X86::VSQRTPDr, X86::VSQRTPDm, TB_ALIGN_16 },
531 { X86::VSQRTPDr_Int, X86::VSQRTPDm_Int, TB_ALIGN_16 },
532 { X86::VSQRTPSr, X86::VSQRTPSm, TB_ALIGN_16 },
533 { X86::VSQRTPSr_Int, X86::VSQRTPSm_Int, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]534 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]535 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]536 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
537
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]538 // AVX 256-bit foldable instructions
539 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
540 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]541 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]542 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]543 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]544 { X86::VPERMILPDYri, X86::VPERMILPDYmi, TB_ALIGN_32 },
545 { X86::VPERMILPSYri, X86::VPERMILPSYmi, TB_ALIGN_32 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]546
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]547 // AVX2 foldable instructions
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]548 { X86::VPABSBrr256, X86::VPABSBrm256, TB_ALIGN_32 },
549 { X86::VPABSDrr256, X86::VPABSDrm256, TB_ALIGN_32 },
550 { X86::VPABSWrr256, X86::VPABSWrm256, TB_ALIGN_32 },
551 { X86::VPSHUFDYri, X86::VPSHUFDYmi, TB_ALIGN_32 },
552 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, TB_ALIGN_32 },
553 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, TB_ALIGN_32 },
554 { X86::VRCPPSYr, X86::VRCPPSYm, TB_ALIGN_32 },
555 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, TB_ALIGN_32 },
556 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, TB_ALIGN_32 },
557 { X86::VRSQRTPSYr_Int, X86::VRSQRTPSYm_Int, TB_ALIGN_32 },
558 { X86::VSQRTPDYr, X86::VSQRTPDYm, TB_ALIGN_32 },
559 { X86::VSQRTPDYr_Int, X86::VSQRTPDYm_Int, TB_ALIGN_32 },
560 { X86::VSQRTPSYr, X86::VSQRTPSYm, TB_ALIGN_32 },
561 { X86::VSQRTPSYr_Int, X86::VSQRTPSYm_Int, TB_ALIGN_32 },
Nadav Rotemee3552f2012-07-15 12:26:30 +0000[diff] [blame]562 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
563 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
Michael Liao2de86af2012-09-26 08:24:51 +0000[diff] [blame^]564
565 // BMI/BMI2 foldable instructions
566 { X86::RORX32ri, X86::RORX32mi, 0 },
567 { X86::RORX64ri, X86::RORX64mi, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]568 };
569
570 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]571 unsigned RegOp = OpTbl1[i].RegOp;
572 unsigned MemOp = OpTbl1[i].MemOp;
573 unsigned Flags = OpTbl1[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]574 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
575 RegOp, MemOp,
576 // Index 1, folded load
577 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]578 }
579
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]580 static const X86OpTblEntry OpTbl2[] = {
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]581 { X86::ADC32rr, X86::ADC32rm, 0 },
582 { X86::ADC64rr, X86::ADC64rm, 0 },
583 { X86::ADD16rr, X86::ADD16rm, 0 },
584 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
585 { X86::ADD32rr, X86::ADD32rm, 0 },
586 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
587 { X86::ADD64rr, X86::ADD64rm, 0 },
588 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
589 { X86::ADD8rr, X86::ADD8rm, 0 },
590 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
591 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
592 { X86::ADDSDrr, X86::ADDSDrm, 0 },
593 { X86::ADDSSrr, X86::ADDSSrm, 0 },
594 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
595 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
596 { X86::AND16rr, X86::AND16rm, 0 },
597 { X86::AND32rr, X86::AND32rm, 0 },
598 { X86::AND64rr, X86::AND64rm, 0 },
599 { X86::AND8rr, X86::AND8rm, 0 },
600 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
601 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
602 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
603 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]604 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
605 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
606 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
607 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]608 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
609 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
610 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
611 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
612 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
613 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
614 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
615 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
616 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
617 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
618 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
619 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
620 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
621 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
622 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
623 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
624 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
625 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
626 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
627 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
628 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
629 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
630 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
631 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
632 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
633 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
634 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
635 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
636 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
637 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
638 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
639 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
640 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
641 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
642 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
643 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
644 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
645 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
646 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
647 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
648 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
649 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
650 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
651 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
652 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
653 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
654 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
655 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
656 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
657 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
658 { X86::CMPSDrr, X86::CMPSDrm, 0 },
659 { X86::CMPSSrr, X86::CMPSSrm, 0 },
660 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
661 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
662 { X86::DIVSDrr, X86::DIVSDrm, 0 },
663 { X86::DIVSSrr, X86::DIVSSrm, 0 },
664 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
665 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
666 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
667 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
668 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
669 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
670 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
671 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
672 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
673 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
674 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
675 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
676 { X86::IMUL16rr, X86::IMUL16rm, 0 },
677 { X86::IMUL32rr, X86::IMUL32rm, 0 },
678 { X86::IMUL64rr, X86::IMUL64rm, 0 },
679 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
680 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
Manman Ren959acb12012-08-13 18:29:41 +0000[diff] [blame]681 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
682 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
683 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
684 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
685 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
686 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]687 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
688 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, TB_ALIGN_16 },
689 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
690 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, TB_ALIGN_16 },
691 { X86::MAXSDrr, X86::MAXSDrm, 0 },
692 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
693 { X86::MAXSSrr, X86::MAXSSrm, 0 },
694 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
695 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
696 { X86::MINPDrr_Int, X86::MINPDrm_Int, TB_ALIGN_16 },
697 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
698 { X86::MINPSrr_Int, X86::MINPSrm_Int, TB_ALIGN_16 },
699 { X86::MINSDrr, X86::MINSDrm, 0 },
700 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
701 { X86::MINSSrr, X86::MINSSrm, 0 },
702 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]703 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]704 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
705 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
706 { X86::MULSDrr, X86::MULSDrm, 0 },
707 { X86::MULSSrr, X86::MULSSrm, 0 },
708 { X86::OR16rr, X86::OR16rm, 0 },
709 { X86::OR32rr, X86::OR32rm, 0 },
710 { X86::OR64rr, X86::OR64rm, 0 },
711 { X86::OR8rr, X86::OR8rm, 0 },
712 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
713 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
714 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
715 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]716 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]717 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
718 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
719 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
720 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
721 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
722 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]723 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
724 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]725 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]726 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]727 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
728 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
729 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
730 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]731 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]732 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
733 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]734 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]735 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
736 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
737 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]738 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]739 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]740 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
741 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]742 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]743 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]744 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]745 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]746 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]747 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]748 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
749 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
750 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
751 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
752 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
753 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]754 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]755 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
756 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
757 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
758 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
759 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
760 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
761 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]762 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
763 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
764 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
765 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]766 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
767 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
768 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
769 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
770 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
771 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
772 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
773 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
774 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
775 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
776 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
777 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
778 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
779 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
780 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
781 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
782 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
783 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
784 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
785 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
786 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
787 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
788 { X86::SBB32rr, X86::SBB32rm, 0 },
789 { X86::SBB64rr, X86::SBB64rm, 0 },
790 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
791 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
792 { X86::SUB16rr, X86::SUB16rm, 0 },
793 { X86::SUB32rr, X86::SUB32rm, 0 },
794 { X86::SUB64rr, X86::SUB64rm, 0 },
795 { X86::SUB8rr, X86::SUB8rm, 0 },
796 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
797 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
798 { X86::SUBSDrr, X86::SUBSDrm, 0 },
799 { X86::SUBSSrr, X86::SUBSSrm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]800 // FIXME: TEST*rr -> swapped operand of TEST*mr.
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]801 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
802 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
803 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
804 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
805 { X86::XOR16rr, X86::XOR16rm, 0 },
806 { X86::XOR32rr, X86::XOR32rm, 0 },
807 { X86::XOR64rr, X86::XOR64rm, 0 },
808 { X86::XOR8rr, X86::XOR8rm, 0 },
809 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]810 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
811 // AVX 128-bit versions of foldable instructions
812 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
813 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
814 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
815 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
816 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
817 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
818 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
819 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
820 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
821 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
822 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
823 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
Craig Topperb6eb5132012-06-25 06:16:00 +0000[diff] [blame]824 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]825 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, TB_ALIGN_16 },
826 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
827 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
828 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
829 { X86::VADDPDrr, X86::VADDPDrm, TB_ALIGN_16 },
830 { X86::VADDPSrr, X86::VADDPSrm, TB_ALIGN_16 },
831 { X86::VADDSDrr, X86::VADDSDrm, 0 },
832 { X86::VADDSSrr, X86::VADDSSrm, 0 },
833 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, TB_ALIGN_16 },
834 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, TB_ALIGN_16 },
835 { X86::VANDNPDrr, X86::VANDNPDrm, TB_ALIGN_16 },
836 { X86::VANDNPSrr, X86::VANDNPSrm, TB_ALIGN_16 },
837 { X86::VANDPDrr, X86::VANDPDrm, TB_ALIGN_16 },
838 { X86::VANDPSrr, X86::VANDPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]839 { X86::VBLENDPDrri, X86::VBLENDPDrmi, TB_ALIGN_16 },
840 { X86::VBLENDPSrri, X86::VBLENDPSrmi, TB_ALIGN_16 },
841 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, TB_ALIGN_16 },
842 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]843 { X86::VCMPPDrri, X86::VCMPPDrmi, TB_ALIGN_16 },
844 { X86::VCMPPSrri, X86::VCMPPSrmi, TB_ALIGN_16 },
845 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
846 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
847 { X86::VDIVPDrr, X86::VDIVPDrm, TB_ALIGN_16 },
848 { X86::VDIVPSrr, X86::VDIVPSrm, TB_ALIGN_16 },
849 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
850 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
851 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
852 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
853 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
854 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
855 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
856 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
857 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
858 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
859 { X86::VHADDPDrr, X86::VHADDPDrm, TB_ALIGN_16 },
860 { X86::VHADDPSrr, X86::VHADDPSrm, TB_ALIGN_16 },
861 { X86::VHSUBPDrr, X86::VHSUBPDrm, TB_ALIGN_16 },
862 { X86::VHSUBPSrr, X86::VHSUBPSrm, TB_ALIGN_16 },
863 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
864 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
865 { X86::VMAXPDrr, X86::VMAXPDrm, TB_ALIGN_16 },
866 { X86::VMAXPDrr_Int, X86::VMAXPDrm_Int, TB_ALIGN_16 },
867 { X86::VMAXPSrr, X86::VMAXPSrm, TB_ALIGN_16 },
868 { X86::VMAXPSrr_Int, X86::VMAXPSrm_Int, TB_ALIGN_16 },
869 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
870 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 },
871 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
872 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 },
873 { X86::VMINPDrr, X86::VMINPDrm, TB_ALIGN_16 },
874 { X86::VMINPDrr_Int, X86::VMINPDrm_Int, TB_ALIGN_16 },
875 { X86::VMINPSrr, X86::VMINPSrm, TB_ALIGN_16 },
876 { X86::VMINPSrr_Int, X86::VMINPSrm_Int, TB_ALIGN_16 },
877 { X86::VMINSDrr, X86::VMINSDrm, 0 },
878 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 },
879 { X86::VMINSSrr, X86::VMINSSrm, 0 },
880 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]881 { X86::VMPSADBWrri, X86::VMPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]882 { X86::VMULPDrr, X86::VMULPDrm, TB_ALIGN_16 },
883 { X86::VMULPSrr, X86::VMULPSrm, TB_ALIGN_16 },
884 { X86::VMULSDrr, X86::VMULSDrm, 0 },
885 { X86::VMULSSrr, X86::VMULSSrm, 0 },
886 { X86::VORPDrr, X86::VORPDrm, TB_ALIGN_16 },
887 { X86::VORPSrr, X86::VORPSrm, TB_ALIGN_16 },
888 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, TB_ALIGN_16 },
889 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]890 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]891 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, TB_ALIGN_16 },
892 { X86::VPADDBrr, X86::VPADDBrm, TB_ALIGN_16 },
893 { X86::VPADDDrr, X86::VPADDDrm, TB_ALIGN_16 },
894 { X86::VPADDQrr, X86::VPADDQrm, TB_ALIGN_16 },
895 { X86::VPADDSBrr, X86::VPADDSBrm, TB_ALIGN_16 },
896 { X86::VPADDSWrr, X86::VPADDSWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]897 { X86::VPADDUSBrr, X86::VPADDUSBrm, TB_ALIGN_16 },
898 { X86::VPADDUSWrr, X86::VPADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]899 { X86::VPADDWrr, X86::VPADDWrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]900 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]901 { X86::VPANDNrr, X86::VPANDNrm, TB_ALIGN_16 },
902 { X86::VPANDrr, X86::VPANDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]903 { X86::VPAVGBrr, X86::VPAVGBrm, TB_ALIGN_16 },
904 { X86::VPAVGWrr, X86::VPAVGWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]905 { X86::VPBLENDWrri, X86::VPBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]906 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, TB_ALIGN_16 },
907 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]908 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]909 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, TB_ALIGN_16 },
910 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, TB_ALIGN_16 },
911 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]912 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]913 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]914 { X86::VPHADDDrr, X86::VPHADDDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]915 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]916 { X86::VPHADDWrr, X86::VPHADDWrm, TB_ALIGN_16 },
917 { X86::VPHSUBDrr, X86::VPHSUBDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]918 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, TB_ALIGN_16 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]919 { X86::VPHSUBWrr, X86::VPHSUBWrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]920 { X86::VPERMILPDrr, X86::VPERMILPDrm, TB_ALIGN_16 },
921 { X86::VPERMILPSrr, X86::VPERMILPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]922 { X86::VPINSRWrri, X86::VPINSRWrmi, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]923 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]924 { X86::VPMADDWDrr, X86::VPMADDWDrm, TB_ALIGN_16 },
925 { X86::VPMAXSWrr, X86::VPMAXSWrm, TB_ALIGN_16 },
926 { X86::VPMAXUBrr, X86::VPMAXUBrm, TB_ALIGN_16 },
927 { X86::VPMINSWrr, X86::VPMINSWrm, TB_ALIGN_16 },
928 { X86::VPMINUBrr, X86::VPMINUBrm, TB_ALIGN_16 },
929 { X86::VPMULDQrr, X86::VPMULDQrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]930 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]931 { X86::VPMULHUWrr, X86::VPMULHUWrm, TB_ALIGN_16 },
932 { X86::VPMULHWrr, X86::VPMULHWrm, TB_ALIGN_16 },
933 { X86::VPMULLDrr, X86::VPMULLDrm, TB_ALIGN_16 },
934 { X86::VPMULLWrr, X86::VPMULLWrm, TB_ALIGN_16 },
935 { X86::VPMULUDQrr, X86::VPMULUDQrm, TB_ALIGN_16 },
936 { X86::VPORrr, X86::VPORrm, TB_ALIGN_16 },
937 { X86::VPSADBWrr, X86::VPSADBWrm, TB_ALIGN_16 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]938 { X86::VPSHUFBrr, X86::VPSHUFBrm, TB_ALIGN_16 },
939 { X86::VPSIGNBrr, X86::VPSIGNBrm, TB_ALIGN_16 },
940 { X86::VPSIGNWrr, X86::VPSIGNWrm, TB_ALIGN_16 },
941 { X86::VPSIGNDrr, X86::VPSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]942 { X86::VPSLLDrr, X86::VPSLLDrm, TB_ALIGN_16 },
943 { X86::VPSLLQrr, X86::VPSLLQrm, TB_ALIGN_16 },
944 { X86::VPSLLWrr, X86::VPSLLWrm, TB_ALIGN_16 },
945 { X86::VPSRADrr, X86::VPSRADrm, TB_ALIGN_16 },
946 { X86::VPSRAWrr, X86::VPSRAWrm, TB_ALIGN_16 },
947 { X86::VPSRLDrr, X86::VPSRLDrm, TB_ALIGN_16 },
948 { X86::VPSRLQrr, X86::VPSRLQrm, TB_ALIGN_16 },
949 { X86::VPSRLWrr, X86::VPSRLWrm, TB_ALIGN_16 },
950 { X86::VPSUBBrr, X86::VPSUBBrm, TB_ALIGN_16 },
951 { X86::VPSUBDrr, X86::VPSUBDrm, TB_ALIGN_16 },
952 { X86::VPSUBSBrr, X86::VPSUBSBrm, TB_ALIGN_16 },
953 { X86::VPSUBSWrr, X86::VPSUBSWrm, TB_ALIGN_16 },
954 { X86::VPSUBWrr, X86::VPSUBWrm, TB_ALIGN_16 },
955 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, TB_ALIGN_16 },
956 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, TB_ALIGN_16 },
957 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, TB_ALIGN_16 },
958 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, TB_ALIGN_16 },
959 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, TB_ALIGN_16 },
960 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, TB_ALIGN_16 },
961 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, TB_ALIGN_16 },
962 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, TB_ALIGN_16 },
963 { X86::VPXORrr, X86::VPXORrm, TB_ALIGN_16 },
964 { X86::VSHUFPDrri, X86::VSHUFPDrmi, TB_ALIGN_16 },
965 { X86::VSHUFPSrri, X86::VSHUFPSrmi, TB_ALIGN_16 },
966 { X86::VSUBPDrr, X86::VSUBPDrm, TB_ALIGN_16 },
967 { X86::VSUBPSrr, X86::VSUBPSrm, TB_ALIGN_16 },
968 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
969 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
970 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, TB_ALIGN_16 },
971 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, TB_ALIGN_16 },
972 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, TB_ALIGN_16 },
973 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, TB_ALIGN_16 },
974 { X86::VXORPDrr, X86::VXORPDrm, TB_ALIGN_16 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]975 { X86::VXORPSrr, X86::VXORPSrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]976 // AVX 256-bit foldable instructions
977 { X86::VADDPDYrr, X86::VADDPDYrm, TB_ALIGN_32 },
978 { X86::VADDPSYrr, X86::VADDPSYrm, TB_ALIGN_32 },
979 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, TB_ALIGN_32 },
980 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, TB_ALIGN_32 },
981 { X86::VANDNPDYrr, X86::VANDNPDYrm, TB_ALIGN_32 },
982 { X86::VANDNPSYrr, X86::VANDNPSYrm, TB_ALIGN_32 },
983 { X86::VANDPDYrr, X86::VANDPDYrm, TB_ALIGN_32 },
984 { X86::VANDPSYrr, X86::VANDPSYrm, TB_ALIGN_32 },
985 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, TB_ALIGN_32 },
986 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, TB_ALIGN_32 },
987 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, TB_ALIGN_32 },
988 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, TB_ALIGN_32 },
989 { X86::VCMPPDYrri, X86::VCMPPDYrmi, TB_ALIGN_32 },
990 { X86::VCMPPSYrri, X86::VCMPPSYrmi, TB_ALIGN_32 },
991 { X86::VDIVPDYrr, X86::VDIVPDYrm, TB_ALIGN_32 },
992 { X86::VDIVPSYrr, X86::VDIVPSYrm, TB_ALIGN_32 },
993 { X86::VHADDPDYrr, X86::VHADDPDYrm, TB_ALIGN_32 },
994 { X86::VHADDPSYrr, X86::VHADDPSYrm, TB_ALIGN_32 },
995 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, TB_ALIGN_32 },
996 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, TB_ALIGN_32 },
997 { X86::VINSERTF128rr, X86::VINSERTF128rm, TB_ALIGN_32 },
998 { X86::VMAXPDYrr, X86::VMAXPDYrm, TB_ALIGN_32 },
999 { X86::VMAXPDYrr_Int, X86::VMAXPDYrm_Int, TB_ALIGN_32 },
1000 { X86::VMAXPSYrr, X86::VMAXPSYrm, TB_ALIGN_32 },
1001 { X86::VMAXPSYrr_Int, X86::VMAXPSYrm_Int, TB_ALIGN_32 },
1002 { X86::VMINPDYrr, X86::VMINPDYrm, TB_ALIGN_32 },
1003 { X86::VMINPDYrr_Int, X86::VMINPDYrm_Int, TB_ALIGN_32 },
1004 { X86::VMINPSYrr, X86::VMINPSYrm, TB_ALIGN_32 },
1005 { X86::VMINPSYrr_Int, X86::VMINPSYrm_Int, TB_ALIGN_32 },
1006 { X86::VMULPDYrr, X86::VMULPDYrm, TB_ALIGN_32 },
1007 { X86::VMULPSYrr, X86::VMULPSYrm, TB_ALIGN_32 },
1008 { X86::VORPDYrr, X86::VORPDYrm, TB_ALIGN_32 },
1009 { X86::VORPSYrr, X86::VORPSYrm, TB_ALIGN_32 },
1010 { X86::VPERM2F128rr, X86::VPERM2F128rm, TB_ALIGN_32 },
1011 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, TB_ALIGN_32 },
1012 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, TB_ALIGN_32 },
1013 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, TB_ALIGN_32 },
1014 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, TB_ALIGN_32 },
1015 { X86::VSUBPDYrr, X86::VSUBPDYrm, TB_ALIGN_32 },
1016 { X86::VSUBPSYrr, X86::VSUBPSYrm, TB_ALIGN_32 },
1017 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, TB_ALIGN_32 },
1018 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, TB_ALIGN_32 },
1019 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, TB_ALIGN_32 },
1020 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, TB_ALIGN_32 },
1021 { X86::VXORPDYrr, X86::VXORPDYrm, TB_ALIGN_32 },
1022 { X86::VXORPSYrr, X86::VXORPSYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1023 // AVX2 foldable instructions
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1024 { X86::VINSERTI128rr, X86::VINSERTI128rm, TB_ALIGN_16 },
1025 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, TB_ALIGN_32 },
1026 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, TB_ALIGN_32 },
1027 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, TB_ALIGN_32 },
1028 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, TB_ALIGN_32 },
1029 { X86::VPADDBYrr, X86::VPADDBYrm, TB_ALIGN_32 },
1030 { X86::VPADDDYrr, X86::VPADDDYrm, TB_ALIGN_32 },
1031 { X86::VPADDQYrr, X86::VPADDQYrm, TB_ALIGN_32 },
1032 { X86::VPADDSBYrr, X86::VPADDSBYrm, TB_ALIGN_32 },
1033 { X86::VPADDSWYrr, X86::VPADDSWYrm, TB_ALIGN_32 },
1034 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, TB_ALIGN_32 },
1035 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, TB_ALIGN_32 },
1036 { X86::VPADDWYrr, X86::VPADDWYrm, TB_ALIGN_32 },
1037 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, TB_ALIGN_32 },
1038 { X86::VPANDNYrr, X86::VPANDNYrm, TB_ALIGN_32 },
1039 { X86::VPANDYrr, X86::VPANDYrm, TB_ALIGN_32 },
1040 { X86::VPAVGBYrr, X86::VPAVGBYrm, TB_ALIGN_32 },
1041 { X86::VPAVGWYrr, X86::VPAVGWYrm, TB_ALIGN_32 },
1042 { X86::VPBLENDDrri, X86::VPBLENDDrmi, TB_ALIGN_32 },
1043 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, TB_ALIGN_32 },
1044 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, TB_ALIGN_32 },
1045 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, TB_ALIGN_32 },
1046 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, TB_ALIGN_32 },
1047 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, TB_ALIGN_32 },
1048 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, TB_ALIGN_32 },
1049 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, TB_ALIGN_32 },
1050 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, TB_ALIGN_32 },
1051 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, TB_ALIGN_32 },
1052 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, TB_ALIGN_32 },
1053 { X86::VPERM2I128rr, X86::VPERM2I128rm, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]1054 { X86::VPERMDYrr, X86::VPERMDYrm, TB_ALIGN_32 },
Elena Demikhovsky779a72b2012-04-15 11:18:59 +0000[diff] [blame]1055 { X86::VPERMPDYri, X86::VPERMPDYmi, TB_ALIGN_32 },
Craig Toppera875b7c2012-01-19 08:50:38 +0000[diff] [blame]1056 { X86::VPERMPSYrr, X86::VPERMPSYrm, TB_ALIGN_32 },
Elena Demikhovsky779a72b2012-04-15 11:18:59 +0000[diff] [blame]1057 { X86::VPERMQYri, X86::VPERMQYmi, TB_ALIGN_32 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]1058 { X86::VPHADDDYrr, X86::VPHADDDYrm, TB_ALIGN_32 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1059 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, TB_ALIGN_32 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]1060 { X86::VPHADDWYrr, X86::VPHADDWYrm, TB_ALIGN_32 },
1061 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, TB_ALIGN_32 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1062 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, TB_ALIGN_32 },
Craig Topperce4f9c52012-01-25 05:37:32 +0000[diff] [blame]1063 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, TB_ALIGN_32 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1064 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, TB_ALIGN_32 },
1065 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, TB_ALIGN_32 },
1066 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, TB_ALIGN_32 },
1067 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, TB_ALIGN_32 },
1068 { X86::VPMINSWYrr, X86::VPMINSWYrm, TB_ALIGN_32 },
1069 { X86::VPMINUBYrr, X86::VPMINUBYrm, TB_ALIGN_32 },
1070 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, TB_ALIGN_32 },
1071 { X86::VPMULDQYrr, X86::VPMULDQYrm, TB_ALIGN_32 },
1072 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, TB_ALIGN_32 },
1073 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, TB_ALIGN_32 },
1074 { X86::VPMULHWYrr, X86::VPMULHWYrm, TB_ALIGN_32 },
1075 { X86::VPMULLDYrr, X86::VPMULLDYrm, TB_ALIGN_32 },
1076 { X86::VPMULLWYrr, X86::VPMULLWYrm, TB_ALIGN_32 },
1077 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, TB_ALIGN_32 },
1078 { X86::VPORYrr, X86::VPORYrm, TB_ALIGN_32 },
1079 { X86::VPSADBWYrr, X86::VPSADBWYrm, TB_ALIGN_32 },
Craig Topper78349002012-01-25 06:43:11 +0000[diff] [blame]1080 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, TB_ALIGN_32 },
1081 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, TB_ALIGN_32 },
1082 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, TB_ALIGN_32 },
1083 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1084 { X86::VPSLLDYrr, X86::VPSLLDYrm, TB_ALIGN_16 },
1085 { X86::VPSLLQYrr, X86::VPSLLQYrm, TB_ALIGN_16 },
1086 { X86::VPSLLWYrr, X86::VPSLLWYrm, TB_ALIGN_16 },
1087 { X86::VPSLLVDrr, X86::VPSLLVDrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1088 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1089 { X86::VPSLLVQrr, X86::VPSLLVQrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1090 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1091 { X86::VPSRADYrr, X86::VPSRADYrm, TB_ALIGN_16 },
1092 { X86::VPSRAWYrr, X86::VPSRAWYrm, TB_ALIGN_16 },
1093 { X86::VPSRAVDrr, X86::VPSRAVDrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1094 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1095 { X86::VPSRLDYrr, X86::VPSRLDYrm, TB_ALIGN_16 },
1096 { X86::VPSRLQYrr, X86::VPSRLQYrm, TB_ALIGN_16 },
1097 { X86::VPSRLWYrr, X86::VPSRLWYrm, TB_ALIGN_16 },
1098 { X86::VPSRLVDrr, X86::VPSRLVDrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1099 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1100 { X86::VPSRLVQrr, X86::VPSRLVQrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1101 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, TB_ALIGN_32 },
1102 { X86::VPSUBBYrr, X86::VPSUBBYrm, TB_ALIGN_32 },
1103 { X86::VPSUBDYrr, X86::VPSUBDYrm, TB_ALIGN_32 },
1104 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, TB_ALIGN_32 },
1105 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, TB_ALIGN_32 },
1106 { X86::VPSUBWYrr, X86::VPSUBWYrm, TB_ALIGN_32 },
1107 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, TB_ALIGN_32 },
1108 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, TB_ALIGN_32 },
Craig Topper182b00a2011-11-14 08:07:55 +0000[diff] [blame]1109 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, TB_ALIGN_16 },
Craig Topperd78429f2012-01-14 18:14:53 +0000[diff] [blame]1110 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, TB_ALIGN_32 },
1111 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, TB_ALIGN_32 },
1112 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, TB_ALIGN_32 },
1113 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, TB_ALIGN_32 },
1114 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, TB_ALIGN_32 },
1115 { X86::VPXORYrr, X86::VPXORYrm, TB_ALIGN_32 },
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1116 // FIXME: add AVX 256-bit foldable instructions
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1117
1118 // FMA4 foldable patterns
1119 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, TB_ALIGN_16 },
1120 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, TB_ALIGN_16 },
1121 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1122 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1123 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1124 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
1125 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1126 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1127 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1128 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
1129 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, TB_ALIGN_16 },
1130 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, TB_ALIGN_16 },
1131 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1132 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1133 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1134 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
1135 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1136 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1137 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1138 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1139 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1140 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1141 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1142 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1143 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1144 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1145 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1146 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
Michael Liaof9f7b552012-09-26 08:22:37 +0000[diff] [blame]1147
1148 // BMI/BMI2 foldable instructions
1149 { X86::MULX32rr, X86::MULX32rm, 0 },
1150 { X86::MULX64rr, X86::MULX64rm, 0 },
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1151 };
1152
1153 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]1154 unsigned RegOp = OpTbl2[i].RegOp;
1155 unsigned MemOp = OpTbl2[i].MemOp;
1156 unsigned Flags = OpTbl2[i].Flags;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1157 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1158 RegOp, MemOp,
1159 // Index 2, folded load
1160 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]1161 }
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1162
1163 static const X86OpTblEntry OpTbl3[] = {
1164 // FMA foldable instructions
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1165 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1166 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1167 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1168 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1169 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1170 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1171 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, 0 },
1172 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1173
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1174 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1175 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1176 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1177 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1178 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1179 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1180 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1181 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1182 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1183 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1184 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1185 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1186
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1187 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1188 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1189 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1190 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1191 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1192 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1193 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, 0 },
1194 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1195
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1196 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1197 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1198 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1199 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1200 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1201 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1202 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1203 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1204 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1205 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1206 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1207 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1208
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1209 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1210 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1211 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1212 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1213 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1214 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1215 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, 0 },
1216 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, 0 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1217
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1218 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1219 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1220 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1221 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1222 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1223 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1224 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1225 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1226 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1227 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1228 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1229 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1230
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1231 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1232 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1233 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1234 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1235 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1236 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]1237 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, 0 },
1238 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, 0 },
Craig Topper2e127b52012-06-01 05:48:39 +0000[diff] [blame]1239
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1240 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1241 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1242 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1243 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1244 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1245 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1246 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1247 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1248 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1249 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1250 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1251 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1252
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1253 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1254 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1255 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1256 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1257 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1258 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1259 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1260 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1261 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1262 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1263 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1264 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
Craig Topper3cb14302012-06-04 07:08:21 +0000[diff] [blame]1265
Craig Topperc6ac4ce2012-06-04 07:46:16 +0000[diff] [blame]1266 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1267 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1268 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1269 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1270 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1271 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1272 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1273 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1274 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1275 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1276 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1277 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
Craig Topper908e6852012-08-31 23:10:34 +0000[diff] [blame]1278
1279 // FMA4 foldable patterns
1280 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, TB_ALIGN_16 },
1281 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, TB_ALIGN_16 },
1282 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1283 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1284 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1285 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
1286 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1287 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1288 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1289 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
1290 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, TB_ALIGN_16 },
1291 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, TB_ALIGN_16 },
1292 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1293 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1294 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1295 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
1296 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1297 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1298 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1299 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1300 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1301 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1302 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1303 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1304 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1305 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1306 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1307 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
Elena Demikhovsky602f3a22012-05-31 09:20:20 +0000[diff] [blame]1308 };
1309
1310 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1311 unsigned RegOp = OpTbl3[i].RegOp;
1312 unsigned MemOp = OpTbl3[i].MemOp;
1313 unsigned Flags = OpTbl3[i].Flags;
1314 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1315 RegOp, MemOp,
1316 // Index 3, folded load
1317 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1318 }
1319
Chris Lattnerd92fb002002-10-25 22:55:53 +0000[diff] [blame]1320}
1321
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]1322void
1323X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1324 MemOp2RegOpTableType &M2RTable,
1325 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1326 if ((Flags & TB_NO_FORWARD) == 0) {
1327 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1328 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1329 }
1330 if ((Flags & TB_NO_REVERSE) == 0) {
1331 assert(!M2RTable.count(MemOp) &&
1332 "Duplicated entries in unfolding maps?");
1333 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1334 }
1335}
1336
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1337bool
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1338X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1339 unsigned &SrcReg, unsigned &DstReg,
1340 unsigned &SubIdx) const {
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1341 switch (MI.getOpcode()) {
1342 default: break;
1343 case X86::MOVSX16rr8:
1344 case X86::MOVZX16rr8:
1345 case X86::MOVSX32rr8:
1346 case X86::MOVZX32rr8:
1347 case X86::MOVSX64rr8:
1348 case X86::MOVZX64rr8:
Evan Chengceb5a4e2010-01-13 08:01:32 +0000[diff] [blame]1349 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1350 // It's not always legal to reference the low 8-bit of the larger
1351 // register in 32-bit mode.
1352 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1353 case X86::MOVSX32rr16:
1354 case X86::MOVZX32rr16:
1355 case X86::MOVSX64rr16:
1356 case X86::MOVZX64rr16:
1357 case X86::MOVSX64rr32:
1358 case X86::MOVZX64rr32: {
1359 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1360 // Be conservative.
1361 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1362 SrcReg = MI.getOperand(1).getReg();
1363 DstReg = MI.getOperand(0).getReg();
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1364 switch (MI.getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1365 default: llvm_unreachable("Unreachable!");
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1366 case X86::MOVSX16rr8:
1367 case X86::MOVZX16rr8:
1368 case X86::MOVSX32rr8:
1369 case X86::MOVZX32rr8:
1370 case X86::MOVSX64rr8:
1371 case X86::MOVZX64rr8:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1372 SubIdx = X86::sub_8bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1373 break;
1374 case X86::MOVSX32rr16:
1375 case X86::MOVZX32rr16:
1376 case X86::MOVSX64rr16:
1377 case X86::MOVZX64rr16:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1378 SubIdx = X86::sub_16bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1379 break;
1380 case X86::MOVSX64rr32:
1381 case X86::MOVZX64rr32:
Jakob Stoklund Olesen396c8802010-05-25 17:04:16 +0000[diff] [blame]1382 SubIdx = X86::sub_32bit;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1383 break;
1384 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1385 return true;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1386 }
1387 }
Evan Cheng30bebff2010-01-13 00:30:23 +0000[diff] [blame]1388 return false;
Evan Cheng42166152010-01-12 00:09:37 +0000[diff] [blame]1389}
1390
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1391/// isFrameOperand - Return true and the FrameIndex if the specified
1392/// operand and follow operands form a reference to the stack frame.
1393bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1394 int &FrameIndex) const {
1395 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1396 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1397 MI->getOperand(Op+1).getImm() == 1 &&
1398 MI->getOperand(Op+2).getReg() == 0 &&
1399 MI->getOperand(Op+3).getImm() == 0) {
1400 FrameIndex = MI->getOperand(Op).getIndex();
1401 return true;
1402 }
1403 return false;
1404}
1405
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1406static bool isFrameLoadOpcode(int Opcode) {
1407 switch (Opcode) {
David Blaikie46a9f012012-01-20 21:51:11 +0000[diff] [blame]1408 default:
1409 return false;
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1410 case X86::MOV8rm:
1411 case X86::MOV16rm:
1412 case X86::MOV32rm:
Evan Cheng11b0a5d2006-09-08 06:48:29 +0000[diff] [blame]1413 case X86::MOV64rm:
Dale Johannesen3d7008c2007-07-04 21:07:47 +0000[diff] [blame]1414 case X86::LD_Fp64m:
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1415 case X86::MOVSSrm:
1416 case X86::MOVSDrm:
Chris Lattnerbfc2c682006-04-18 16:44:51 +0000[diff] [blame]1417 case X86::MOVAPSrm:
1418 case X86::MOVAPDrm:
Dan Gohmanbdc0f8b2009-01-09 02:40:34 +0000[diff] [blame]1419 case X86::MOVDQArm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1420 case X86::VMOVSSrm:
1421 case X86::VMOVSDrm:
1422 case X86::VMOVAPSrm:
1423 case X86::VMOVAPDrm:
1424 case X86::VMOVDQArm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1425 case X86::VMOVAPSYrm:
1426 case X86::VMOVAPDYrm:
1427 case X86::VMOVDQAYrm:
Bill Wendlinge7b2a862007-04-03 06:00:37 +0000[diff] [blame]1428 case X86::MMX_MOVD64rm:
1429 case X86::MMX_MOVQ64rm:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1430 return true;
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1431 }
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1432}
1433
1434static bool isFrameStoreOpcode(int Opcode) {
1435 switch (Opcode) {
1436 default: break;
1437 case X86::MOV8mr:
1438 case X86::MOV16mr:
1439 case X86::MOV32mr:
1440 case X86::MOV64mr:
1441 case X86::ST_FpP64m:
1442 case X86::MOVSSmr:
1443 case X86::MOVSDmr:
1444 case X86::MOVAPSmr:
1445 case X86::MOVAPDmr:
1446 case X86::MOVDQAmr:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]1447 case X86::VMOVSSmr:
1448 case X86::VMOVSDmr:
1449 case X86::VMOVAPSmr:
1450 case X86::VMOVAPDmr:
1451 case X86::VMOVDQAmr:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]1452 case X86::VMOVAPSYmr:
1453 case X86::VMOVAPDYmr:
1454 case X86::VMOVDQAYmr:
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1455 case X86::MMX_MOVD64mr:
1456 case X86::MMX_MOVQ64mr:
1457 case X86::MMX_MOVNTQmr:
1458 return true;
1459 }
1460 return false;
1461}
1462
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1463unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1464 int &FrameIndex) const {
1465 if (isFrameLoadOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1466 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1467 return MI->getOperand(0).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1468 return 0;
1469}
1470
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1471unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1472 int &FrameIndex) const {
1473 if (isFrameLoadOpcode(MI->getOpcode())) {
1474 unsigned Reg;
1475 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1476 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1477 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1478 const MachineMemOperand *Dummy;
1479 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1480 }
1481 return 0;
1482}
1483
Dan Gohman0b273252008-11-18 19:49:32 +0000[diff] [blame]1484unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1485 int &FrameIndex) const {
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1486 if (isFrameStoreOpcode(MI->getOpcode()))
Jakob Stoklund Olesen96a890a2010-07-27 04:17:01 +0000[diff] [blame]1487 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1488 isFrameOperand(MI, 0, FrameIndex))
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]1489 return MI->getOperand(X86::AddrNumOperands).getReg();
David Greene2f4c3742009-11-13 00:29:53 +0000[diff] [blame]1490 return 0;
1491}
1492
1493unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1494 int &FrameIndex) const {
1495 if (isFrameStoreOpcode(MI->getOpcode())) {
1496 unsigned Reg;
1497 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1498 return Reg;
David Greene70fdd572009-11-12 20:55:29 +0000[diff] [blame]1499 // Check for post-frame index elimination operations
David Greene0508e432009-12-04 22:38:46 +0000[diff] [blame]1500 const MachineMemOperand *Dummy;
1501 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
Chris Lattnerbb53acd2006-02-02 20:12:32 +0000[diff] [blame]1502 }
1503 return 0;
1504}
1505
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1506/// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1507/// X86::MOVPC32r.
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1508static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
Jakob Stoklund Olesen3b9a4422012-08-08 00:40:47 +0000[diff] [blame]1509 // Don't waste compile time scanning use-def chains of physregs.
1510 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1511 return false;
Evan Cheng308e5642008-03-27 01:45:11 +0000[diff] [blame]1512 bool isPICBase = false;
1513 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1514 E = MRI.def_end(); I != E; ++I) {
1515 MachineInstr *DefMI = I.getOperand().getParent();
1516 if (DefMI->getOpcode() != X86::MOVPC32r)
1517 return false;
1518 assert(!isPICBase && "More than one PIC base?");
1519 isPICBase = true;
1520 }
1521 return isPICBase;
1522}
Evan Cheng1973a462008-03-31 07:54:19 +0000[diff] [blame]1523
Bill Wendling1e117682008-05-12 20:54:26 +0000[diff] [blame]1524bool
Dan Gohmane919de52009-10-10 00:34:18 +0000[diff] [blame]1525X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1526 AliasAnalysis *AA) const {
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1527 switch (MI->getOpcode()) {
1528 default: break;
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1529 case X86::MOV8rm:
1530 case X86::MOV16rm:
1531 case X86::MOV32rm:
1532 case X86::MOV64rm:
1533 case X86::LD_Fp64m:
1534 case X86::MOVSSrm:
1535 case X86::MOVSDrm:
1536 case X86::MOVAPSrm:
1537 case X86::MOVUPSrm:
1538 case X86::MOVAPDrm:
1539 case X86::MOVDQArm:
1540 case X86::VMOVSSrm:
1541 case X86::VMOVSDrm:
1542 case X86::VMOVAPSrm:
1543 case X86::VMOVUPSrm:
1544 case X86::VMOVAPDrm:
1545 case X86::VMOVDQArm:
1546 case X86::VMOVAPSYrm:
1547 case X86::VMOVUPSYrm:
1548 case X86::VMOVAPDYrm:
1549 case X86::VMOVDQAYrm:
1550 case X86::MMX_MOVD64rm:
1551 case X86::MMX_MOVQ64rm:
1552 case X86::FsVMOVAPSrm:
1553 case X86::FsVMOVAPDrm:
1554 case X86::FsMOVAPSrm:
1555 case X86::FsMOVAPDrm: {
1556 // Loads from constant pools are trivially rematerializable.
1557 if (MI->getOperand(1).isReg() &&
1558 MI->getOperand(2).isImm() &&
1559 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1560 MI->isInvariantLoad(AA)) {
1561 unsigned BaseReg = MI->getOperand(1).getReg();
1562 if (BaseReg == 0 || BaseReg == X86::RIP)
1563 return true;
1564 // Allow re-materialization of PIC load.
1565 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1566 return false;
1567 const MachineFunction &MF = *MI->getParent()->getParent();
1568 const MachineRegisterInfo &MRI = MF.getRegInfo();
1569 return regIsPICBase(BaseReg, MRI);
Evan Cheng94ba37f2008-02-22 09:25:47 +0000[diff] [blame]1570 }
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1571 return false;
1572 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1573
Craig Toppera0cabf12012-08-21 08:17:07 +0000[diff] [blame]1574 case X86::LEA32r:
1575 case X86::LEA64r: {
1576 if (MI->getOperand(2).isImm() &&
1577 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1578 !MI->getOperand(4).isReg()) {
1579 // lea fi#, lea GV, etc. are all rematerializable.
1580 if (!MI->getOperand(1).isReg())
1581 return true;
1582 unsigned BaseReg = MI->getOperand(1).getReg();
1583 if (BaseReg == 0)
1584 return true;
1585 // Allow re-materialization of lea PICBase + x.
1586 const MachineFunction &MF = *MI->getParent()->getParent();
1587 const MachineRegisterInfo &MRI = MF.getRegInfo();
1588 return regIsPICBase(BaseReg, MRI);
1589 }
1590 return false;
1591 }
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1592 }
Evan Cheng29e62a52008-03-27 01:41:09 +0000[diff] [blame]1593
Dan Gohmane8c1e422007-06-26 00:48:07 +0000[diff] [blame]1594 // All other instructions marked M_REMATERIALIZABLE are always trivially
1595 // rematerializable.
1596 return true;
Dan Gohman4a4a8eb2007-06-14 20:50:44 +0000[diff] [blame]1597}
1598
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1599/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1600/// would clobber the EFLAGS condition register. Note the result may be
1601/// conservative. If it cannot definitely determine the safety after visiting
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1602/// a few instructions in each direction it assumes it's not safe.
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1603static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1604 MachineBasicBlock::iterator I) {
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1605 MachineBasicBlock::iterator E = MBB.end();
1606
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1607 // For compile time consideration, if we are not able to determine the
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1608 // safety after visiting 4 instructions in each direction, we will assume
1609 // it's not safe.
1610 MachineBasicBlock::iterator Iter = I;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1611 for (unsigned i = 0; Iter != E && i < 4; ++i) {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1612 bool SeenDef = false;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1613 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1614 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1615 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1616 SeenDef = true;
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1617 if (!MO.isReg())
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1618 continue;
1619 if (MO.getReg() == X86::EFLAGS) {
1620 if (MO.isUse())
1621 return false;
1622 SeenDef = true;
1623 }
1624 }
1625
1626 if (SeenDef)
1627 // This instruction defines EFLAGS, no need to look any further.
1628 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1629 ++Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1630 // Skip over DBG_VALUE.
1631 while (Iter != E && Iter->isDebugValue())
1632 ++Iter;
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1633 }
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1634
Jakob Stoklund Olesenf08354d2011-09-02 23:52:52 +0000[diff] [blame]1635 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1636 // live in.
1637 if (Iter == E) {
1638 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1639 SE = MBB.succ_end(); SI != SE; ++SI)
1640 if ((*SI)->isLiveIn(X86::EFLAGS))
1641 return false;
1642 return true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1643 }
1644
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1645 MachineBasicBlock::iterator B = MBB.begin();
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1646 Iter = I;
1647 for (unsigned i = 0; i < 4; ++i) {
1648 // If we make it to the beginning of the block, it's safe to clobber
1649 // EFLAGS iff EFLAGS is not live-in.
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1650 if (Iter == B)
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1651 return !MBB.isLiveIn(X86::EFLAGS);
1652
1653 --Iter;
Evan Chengb6dee6e2010-03-23 20:35:45 +0000[diff] [blame]1654 // Skip over DBG_VALUE.
1655 while (Iter != B && Iter->isDebugValue())
1656 --Iter;
1657
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1658 bool SawKill = false;
1659 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1660 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen4519fd02012-02-09 00:17:22 +0000[diff] [blame]1661 // A register mask may clobber EFLAGS, but we should still look for a
1662 // live EFLAGS def.
1663 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1664 SawKill = true;
Dan Gohman0be8c2b2009-10-14 00:08:59 +0000[diff] [blame]1665 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1666 if (MO.isDef()) return MO.isDead();
1667 if (MO.isKill()) SawKill = true;
1668 }
1669 }
1670
1671 if (SawKill)
1672 // This instruction kills EFLAGS and doesn't redefine it, so
1673 // there's no need to look further.
Dan Gohmanc8354582008-10-21 03:24:31 +0000[diff] [blame]1674 return true;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1675 }
1676
1677 // Conservative answer.
1678 return false;
1679}
1680
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1681void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1682 MachineBasicBlock::iterator I,
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1683 unsigned DestReg, unsigned SubIdx,
Evan Cheng6ad7da92009-11-14 02:55:43 +0000[diff] [blame]1684 const MachineInstr *Orig,
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1685 const TargetRegisterInfo &TRI) const {
Dan Gohman90c600d2010-05-07 01:28:10 +0000[diff] [blame]1686 DebugLoc DL = Orig->getDebugLoc();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1687
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1688 // MOV32r0 etc. are implemented with xor which clobbers condition code.
1689 // Re-materialize them as movri instructions to avoid side effects.
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1690 bool Clone = true;
1691 unsigned Opc = Orig->getOpcode();
1692 switch (Opc) {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1693 default: break;
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1694 case X86::MOV8r0:
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]1695 case X86::MOV16r0:
1696 case X86::MOV32r0:
1697 case X86::MOV64r0: {
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1698 if (!isSafeToClobberEFLAGS(MBB, I)) {
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1699 switch (Opc) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1700 default: llvm_unreachable("Unreachable!");
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1701 case X86::MOV8r0: Opc = X86::MOV8ri; break;
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]1702 case X86::MOV16r0: Opc = X86::MOV16ri; break;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1703 case X86::MOV32r0: Opc = X86::MOV32ri; break;
Dan Gohman952f6f92010-02-26 16:49:27 +0000[diff] [blame]1704 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1705 }
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1706 Clone = false;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1707 }
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1708 break;
Evan Cheng3f2ceac2008-06-24 07:10:51 +0000[diff] [blame]1709 }
1710 }
1711
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1712 if (Clone) {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1713 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1714 MBB.insert(I, MI);
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1715 } else {
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1716 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1717 }
Evan Cheng147cb762008-04-16 23:44:44 +0000[diff] [blame]1718
Evan Cheng84517442009-07-16 09:20:10 +0000[diff] [blame]1719 MachineInstr *NewMI = prior(I);
Jakob Stoklund Olesena8ad9772010-06-02 22:47:25 +0000[diff] [blame]1720 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
Evan Chenged6e34f2008-03-31 20:40:39 +0000[diff] [blame]1721}
1722
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1723/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1724/// is not marked dead.
1725static bool hasLiveCondCodeDef(MachineInstr *MI) {
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1726 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1727 MachineOperand &MO = MI->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]1728 if (MO.isReg() && MO.isDef() &&
Evan Chenga8a9c152007-10-05 08:04:01 +0000[diff] [blame]1729 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1730 return true;
1731 }
1732 }
1733 return false;
1734}
1735
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1736/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1737/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1738/// to a 32-bit superregister and then truncating back down to a 16-bit
1739/// subregister.
1740MachineInstr *
1741X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1742 MachineFunction::iterator &MFI,
1743 MachineBasicBlock::iterator &MBBI,
1744 LiveVariables *LV) const {
1745 MachineInstr *MI = MBBI;
1746 unsigned Dest = MI->getOperand(0).getReg();
1747 unsigned Src = MI->getOperand(1).getReg();
1748 bool isDead = MI->getOperand(0).isDead();
1749 bool isKill = MI->getOperand(1).isKill();
1750
1751 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1752 ? X86::LEA64_32r : X86::LEA32r;
1753 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1754 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1755 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1756
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1757 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1758 // well be shifting and then extracting the lower 16-bits.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1759 // This has the potential to cause partial register stall. e.g.
Evan Cheng3974c8d2009-12-12 18:55:26 +0000[diff] [blame]1760 // movw (%rbp,%rcx,2), %dx
1761 // leal -65(%rdx), %esi
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1762 // But testing has shown this *does* help performance in 64-bit mode (at
1763 // least on modern x86 machines).
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1764 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1765 MachineInstr *InsMI =
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1766 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1767 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1768 .addReg(Src, getKillRegState(isKill));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1769
1770 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1771 get(Opc), leaOutReg);
1772 switch (MIOpc) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]1773 default: llvm_unreachable("Unreachable!");
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1774 case X86::SHL16ri: {
1775 unsigned ShAmt = MI->getOperand(2).getImm();
1776 MIB.addReg(0).addImm(1 << ShAmt)
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1777 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1778 break;
1779 }
1780 case X86::INC16r:
1781 case X86::INC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1782 addRegOffset(MIB, leaInReg, true, 1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1783 break;
1784 case X86::DEC16r:
1785 case X86::DEC64_16r:
Chris Lattnerf4693072010-07-08 23:46:44 +0000[diff] [blame]1786 addRegOffset(MIB, leaInReg, true, -1);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1787 break;
1788 case X86::ADD16ri:
1789 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]1790 case X86::ADD16ri_DB:
1791 case X86::ADD16ri8_DB:
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1792 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1793 break;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]1794 case X86::ADD16rr:
1795 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1796 unsigned Src2 = MI->getOperand(2).getReg();
1797 bool isKill2 = MI->getOperand(2).isKill();
1798 unsigned leaInReg2 = 0;
1799 MachineInstr *InsMI2 = 0;
1800 if (Src == Src2) {
1801 // ADD16rr %reg1028<kill>, %reg1028
1802 // just a single insert_subreg.
1803 addRegReg(MIB, leaInReg, true, leaInReg, false);
1804 } else {
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1805 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1806 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1807 // well be shifting and then extracting the lower 16-bits.
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1808 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1809 InsMI2 =
Evan Cheng7fae11b2011-12-14 02:11:42 +0000[diff] [blame]1810 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
Jakob Stoklund Olesena1e883d2010-07-08 16:40:15 +0000[diff] [blame]1811 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1812 .addReg(Src2, getKillRegState(isKill2));
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1813 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1814 }
1815 if (LV && isKill2 && InsMI2)
1816 LV->replaceKillInstruction(Src2, MI, InsMI2);
1817 break;
1818 }
1819 }
1820
1821 MachineInstr *NewMI = MIB;
1822 MachineInstr *ExtMI =
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1823 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1824 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Jakob Stoklund Olesen00264622010-07-08 16:40:22 +0000[diff] [blame]1825 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1826
1827 if (LV) {
1828 // Update live variables
1829 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1830 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1831 if (isKill)
1832 LV->replaceKillInstruction(Src, MI, InsMI);
1833 if (isDead)
1834 LV->replaceKillInstruction(Dest, MI, ExtMI);
1835 }
1836
1837 return ExtMI;
1838}
1839
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1840/// convertToThreeAddress - This method must be implemented by targets that
1841/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1842/// may be able to convert a two-address instruction into a true
1843/// three-address instruction on demand. This allows the X86 target (for
1844/// example) to convert ADD and SHL instructions into LEA instructions if they
1845/// would require register copies due to two-addressness.
1846///
1847/// This method returns a null pointer if the transformation cannot be
1848/// performed, otherwise it returns the new instruction.
1849///
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1850MachineInstr *
1851X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1852 MachineBasicBlock::iterator &MBBI,
Owen Anderson30cc0282008-07-02 23:41:07 +0000[diff] [blame]1853 LiveVariables *LV) const {
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1854 MachineInstr *MI = MBBI;
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]1855 MachineFunction &MF = *MI->getParent()->getParent();
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1856 // All instructions input are two-addr instructions. Get the known operands.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1857 const MachineOperand &Dest = MI->getOperand(0);
1858 const MachineOperand &Src = MI->getOperand(1);
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1859
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]1860 MachineInstr *NewMI = NULL;
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1861 // 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]1862 // we have better subtarget support, enable the 16-bit LEA generation here.
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1863 // 16-bit LEA is also slow on Core2.
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1864 bool DisableLEA16 = true;
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1865 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
Evan Cheng07fc1072006-12-01 21:52:41 +0000[diff] [blame]1866
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1867 unsigned MIOpc = MI->getOpcode();
1868 switch (MIOpc) {
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1869 case X86::SHUFPSrri: {
1870 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1871 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]1872
Evan Chengc8c172e2006-05-30 21:45:53 +0000[diff] [blame]1873 unsigned B = MI->getOperand(1).getReg();
1874 unsigned C = MI->getOperand(2).getReg();
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1875 if (B != C) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1876 unsigned M = MI->getOperand(3).getImm();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1877 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1878 .addOperand(Dest).addOperand(Src).addImm(M);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1879 break;
1880 }
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1881 case X86::SHUFPDrri: {
1882 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1883 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1884
1885 unsigned B = MI->getOperand(1).getReg();
1886 unsigned C = MI->getOperand(2).getReg();
1887 if (B != C) return 0;
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1888 unsigned M = MI->getOperand(3).getImm();
1889
1890 // Convert to PSHUFD mask.
1891 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1892
1893 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1894 .addOperand(Dest).addOperand(Src).addImm(M);
Craig Toppere52d86a2012-01-13 09:21:41 +0000[diff] [blame]1895 break;
1896 }
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1897 case X86::SHL64ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1898 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1899 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1900 // the flags produced by a shift yet, so this is safe.
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1901 unsigned ShAmt = MI->getOperand(2).getImm();
1902 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1903
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1904 // LEA can't handle RSP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1905 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1906 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1907 &X86::GR64_NOSPRegClass))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1908 return 0;
1909
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1910 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1911 .addOperand(Dest)
1912 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattnerbcd38852007-03-28 18:12:31 +0000[diff] [blame]1913 break;
1914 }
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1915 case X86::SHL32ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1916 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1917 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1918 // the flags produced by a shift yet, so this is safe.
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1919 unsigned ShAmt = MI->getOperand(2).getImm();
1920 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1921
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1922 // LEA can't handle ESP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1923 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1924 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1925 &X86::GR32_NOSPRegClass))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1926 return 0;
1927
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1928 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]1929 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1930 .addOperand(Dest)
1931 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1932 break;
1933 }
1934 case X86::SHL16ri: {
Evan Cheng483e1ce2007-09-14 21:48:26 +0000[diff] [blame]1935 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Evan Cheng189df732007-09-06 00:14:41 +0000[diff] [blame]1936 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1937 // the flags produced by a shift yet, so this is safe.
Evan Cheng189df732007-09-06 00:14:41 +0000[diff] [blame]1938 unsigned ShAmt = MI->getOperand(2).getImm();
1939 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]1940
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1941 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1942 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1943 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1944 .addOperand(Dest)
1945 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattner3e1d9172007-03-20 06:08:29 +0000[diff] [blame]1946 break;
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1947 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1948 default: {
1949 // The following opcodes also sets the condition code register(s). Only
1950 // convert them to equivalent lea if the condition code register def's
1951 // are dead!
1952 if (hasLiveCondCodeDef(MI))
1953 return 0;
Evan Cheng66f849b2006-05-30 20:26:50 +0000[diff] [blame]1954
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1955 switch (MIOpc) {
1956 default: return 0;
1957 case X86::INC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]1958 case X86::INC32r:
1959 case X86::INC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1960 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]1961 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1962 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]1963 const TargetRegisterClass *RC = MIOpc == X86::INC64r ?
1964 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1965 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1966
1967 // LEA can't handle RSP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1968 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1969 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1970 return 0;
1971
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1972 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1973 .addOperand(Dest).addOperand(Src), 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1974 break;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]1975 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1976 case X86::INC16r:
1977 case X86::INC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]1978 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]1979 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1980 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1981 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1982 .addOperand(Dest).addOperand(Src), 1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1983 break;
1984 case X86::DEC64r:
Dan Gohmanbeac19e2009-01-06 23:34:46 +0000[diff] [blame]1985 case X86::DEC32r:
1986 case X86::DEC64_32r: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]1987 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Evan Cheng82bc90a2007-10-09 07:14:53 +0000[diff] [blame]1988 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1989 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]1990 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ?
1991 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1992 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1993 // LEA can't handle RSP.
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1994 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1995 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]1996 return 0;
1997
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]1998 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1999 .addOperand(Dest).addOperand(Src), -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2000 break;
2001 }
2002 case X86::DEC16r:
2003 case X86::DEC64_16r:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2004 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2005 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2006 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2007 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2008 .addOperand(Dest).addOperand(Src), -1);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2009 break;
2010 case X86::ADD64rr:
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2011 case X86::ADD64rr_DB:
2012 case X86::ADD32rr:
2013 case X86::ADD32rr_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2014 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2015 unsigned Opc;
Craig Topper760b1342012-02-22 05:59:10 +0000[diff] [blame]2016 const TargetRegisterClass *RC;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2017 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
2018 Opc = X86::LEA64r;
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]2019 RC = &X86::GR64_NOSPRegClass;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2020 } else {
2021 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]2022 RC = &X86::GR32_NOSPRegClass;
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2023 }
2024
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2025
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2026 unsigned Src2 = MI->getOperand(2).getReg();
2027 bool isKill2 = MI->getOperand(2).isKill();
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2028
2029 // LEA can't handle RSP.
2030 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2031 !MF.getRegInfo().constrainRegClass(Src2, RC))
Jakob Stoklund Olesenb19bae42010-10-07 00:07:26 +0000[diff] [blame]2032 return 0;
2033
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2034 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2035 .addOperand(Dest),
2036 Src.getReg(), Src.isKill(), Src2, isKill2);
Nadav Rotem4968e452012-07-16 10:52:25 +0000[diff] [blame]2037
2038 // Preserve undefness of the operands.
2039 bool isUndef = MI->getOperand(1).isUndef();
2040 bool isUndef2 = MI->getOperand(2).isUndef();
2041 NewMI->getOperand(1).setIsUndef(isUndef);
2042 NewMI->getOperand(3).setIsUndef(isUndef2);
2043
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2044 if (LV && isKill2)
2045 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2046 break;
2047 }
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]2048 case X86::ADD16rr:
2049 case X86::ADD16rr_DB: {
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2050 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2051 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2052 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2053 unsigned Src2 = MI->getOperand(2).getReg();
2054 bool isKill2 = MI->getOperand(2).isKill();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2055 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2056 .addOperand(Dest),
2057 Src.getReg(), Src.isKill(), Src2, isKill2);
2058
2059 // Preserve undefness of the operands.
2060 bool isUndef = MI->getOperand(1).isUndef();
2061 bool isUndef2 = MI->getOperand(2).isUndef();
2062 NewMI->getOperand(1).setIsUndef(isUndef);
2063 NewMI->getOperand(3).setIsUndef(isUndef2);
2064
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2065 if (LV && isKill2)
2066 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2067 break;
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2068 }
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2069 case X86::ADD64ri32:
2070 case X86::ADD64ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2071 case X86::ADD64ri32_DB:
2072 case X86::ADD64ri8_DB:
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2073 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2074 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2075 .addOperand(Dest).addOperand(Src),
2076 MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2077 break;
2078 case X86::ADD32ri:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2079 case X86::ADD32ri8:
2080 case X86::ADD32ri_DB:
2081 case X86::ADD32ri8_DB: {
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2082 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2083 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2084 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2085 .addOperand(Dest).addOperand(Src),
2086 MI->getOperand(2).getImm());
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2087 break;
2088 }
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2089 case X86::ADD16ri:
2090 case X86::ADD16ri8:
Chris Lattnerdd774772010-10-08 03:57:25 +0000[diff] [blame]2091 case X86::ADD16ri_DB:
2092 case X86::ADD16ri8_DB:
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2093 if (DisableLEA16)
Evan Cheng26fdd722009-12-12 20:03:14 +0000[diff] [blame]2094 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2095 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2096 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2097 .addOperand(Dest).addOperand(Src),
2098 MI->getOperand(2).getImm());
Evan Cheng766a73f2009-12-11 06:01:48 +0000[diff] [blame]2099 break;
Evan Chengfa2c8282007-10-05 20:34:26 +0000[diff] [blame]2100 }
2101 }
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2102 }
2103
Evan Cheng1bc1cae2008-02-07 08:29:53 +0000[diff] [blame]2104 if (!NewMI) return 0;
2105
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2106 if (LV) { // Update live variables
Jakob Stoklund Olesen70304272012-08-23 22:36:31 +0000[diff] [blame]2107 if (Src.isKill())
2108 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2109 if (Dest.isDead())
2110 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2111 }
2112
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2113 MFI->insert(MBBI, NewMI); // Insert the new inst
Evan Chengdc2c8742006-11-15 20:58:11 +0000[diff] [blame]2114 return NewMI;
Chris Lattnerb7782d72005-01-02 02:37:07 +0000[diff] [blame]2115}
2116
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2117/// commuteInstruction - We have a few instructions that must be hacked on to
2118/// commute them.
2119///
Evan Cheng03553bb2008-06-16 07:33:11 +0000[diff] [blame]2120MachineInstr *
2121X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2122 switch (MI->getOpcode()) {
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2123 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2124 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2125 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2126 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2127 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2128 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2129 unsigned Opc;
2130 unsigned Size;
2131 switch (MI->getOpcode()) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2132 default: llvm_unreachable("Unreachable!");
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2133 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2134 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2135 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2136 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
Dan Gohman48ea03d2007-09-14 23:17:45 +0000[diff] [blame]2137 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2138 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
Chris Lattnerd54845f2005-01-19 07:31:24 +0000[diff] [blame]2139 }
Chris Lattner5c463782007-12-30 20:49:49 +0000[diff] [blame]2140 unsigned Amt = MI->getOperand(3).getImm();
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2141 if (NewMI) {
2142 MachineFunction &MF = *MI->getParent()->getParent();
2143 MI = MF.CloneMachineInstr(MI);
2144 NewMI = false;
Evan Cheng244183e2008-02-13 02:46:49 +0000[diff] [blame]2145 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2146 MI->setDesc(get(Opc));
2147 MI->getOperand(3).setImm(Size-Amt);
2148 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2149 }
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2150 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2151 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2152 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2153 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2154 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2155 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2156 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2157 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2158 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2159 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2160 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2161 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2162 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2163 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2164 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2165 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2166 unsigned Opc;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2167 switch (MI->getOpcode()) {
Craig Topper653e7592012-08-21 07:32:16 +0000[diff] [blame]2168 default: llvm_unreachable("Unreachable!");
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2169 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2170 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2171 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2172 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2173 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2174 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2175 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2176 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2177 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2178 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2179 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2180 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
Chris Lattner1a1c6002010-10-05 23:00:14 +0000[diff] [blame]2181 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2182 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2183 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2184 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2185 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2186 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2187 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2188 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2189 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2190 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2191 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2192 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2193 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2194 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2195 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2196 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2197 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2198 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2199 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2200 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2201 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2202 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2203 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2204 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2205 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2206 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2207 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2208 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2209 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2210 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2211 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2212 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
Mon P Wang6c8bcf92009-04-18 05:16:01 +0000[diff] [blame]2213 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
Dan Gohman7e47cc72009-01-07 00:35:10 +0000[diff] [blame]2214 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2215 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2216 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2217 }
Dan Gohmana39b0a12008-10-17 01:23:35 +0000[diff] [blame]2218 if (NewMI) {
2219 MachineFunction &MF = *MI->getParent()->getParent();
2220 MI = MF.CloneMachineInstr(MI);
2221 NewMI = false;
2222 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]2223 MI->setDesc(get(Opc));
Evan Cheng1151ffd2007-10-05 23:13:21 +0000[diff] [blame]2224 // Fallthrough intended.
2225 }
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2226 default:
Evan Cheng03553bb2008-06-16 07:33:11 +0000[diff] [blame]2227 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner29478012005-01-19 07:11:01 +0000[diff] [blame]2228 }
2229}
2230
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2231static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2232 switch (BrOpc) {
2233 default: return X86::COND_INVALID;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2234 case X86::JE_4: return X86::COND_E;
2235 case X86::JNE_4: return X86::COND_NE;
2236 case X86::JL_4: return X86::COND_L;
2237 case X86::JLE_4: return X86::COND_LE;
2238 case X86::JG_4: return X86::COND_G;
2239 case X86::JGE_4: return X86::COND_GE;
2240 case X86::JB_4: return X86::COND_B;
2241 case X86::JBE_4: return X86::COND_BE;
2242 case X86::JA_4: return X86::COND_A;
2243 case X86::JAE_4: return X86::COND_AE;
2244 case X86::JS_4: return X86::COND_S;
2245 case X86::JNS_4: return X86::COND_NS;
2246 case X86::JP_4: return X86::COND_P;
2247 case X86::JNP_4: return X86::COND_NP;
2248 case X86::JO_4: return X86::COND_O;
2249 case X86::JNO_4: return X86::COND_NO;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2250 }
2251}
2252
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2253/// getCondFromSETOpc - return condition code of a SET opcode.
2254static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2255 switch (Opc) {
2256 default: return X86::COND_INVALID;
2257 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2258 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2259 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2260 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2261 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2262 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2263 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2264 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2265 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2266 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2267 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2268 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2269 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2270 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2271 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2272 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2273 }
2274}
2275
2276/// getCondFromCmovOpc - return condition code of a CMov opcode.
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]2277X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2278 switch (Opc) {
2279 default: return X86::COND_INVALID;
2280 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2281 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2282 return X86::COND_A;
2283 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2284 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2285 return X86::COND_AE;
2286 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2287 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2288 return X86::COND_B;
2289 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2290 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2291 return X86::COND_BE;
2292 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2293 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2294 return X86::COND_E;
2295 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2296 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2297 return X86::COND_G;
2298 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2299 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2300 return X86::COND_GE;
2301 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2302 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2303 return X86::COND_L;
2304 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2305 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2306 return X86::COND_LE;
2307 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2308 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2309 return X86::COND_NE;
2310 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2311 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2312 return X86::COND_NO;
2313 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2314 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2315 return X86::COND_NP;
2316 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2317 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2318 return X86::COND_NS;
2319 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2320 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2321 return X86::COND_O;
2322 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2323 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2324 return X86::COND_P;
2325 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2326 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2327 return X86::COND_S;
2328 }
2329}
2330
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2331unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2332 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2333 default: llvm_unreachable("Illegal condition code!");
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2334 case X86::COND_E: return X86::JE_4;
2335 case X86::COND_NE: return X86::JNE_4;
2336 case X86::COND_L: return X86::JL_4;
2337 case X86::COND_LE: return X86::JLE_4;
2338 case X86::COND_G: return X86::JG_4;
2339 case X86::COND_GE: return X86::JGE_4;
2340 case X86::COND_B: return X86::JB_4;
2341 case X86::COND_BE: return X86::JBE_4;
2342 case X86::COND_A: return X86::JA_4;
2343 case X86::COND_AE: return X86::JAE_4;
2344 case X86::COND_S: return X86::JS_4;
2345 case X86::COND_NS: return X86::JNS_4;
2346 case X86::COND_P: return X86::JP_4;
2347 case X86::COND_NP: return X86::JNP_4;
2348 case X86::COND_O: return X86::JO_4;
2349 case X86::COND_NO: return X86::JNO_4;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2350 }
2351}
2352
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2353/// GetOppositeBranchCondition - Return the inverse of the specified condition,
2354/// e.g. turning COND_E to COND_NE.
2355X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2356 switch (CC) {
Torok Edwinfbcc6632009-07-14 16:55:14 +0000[diff] [blame]2357 default: llvm_unreachable("Illegal condition code!");
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2358 case X86::COND_E: return X86::COND_NE;
2359 case X86::COND_NE: return X86::COND_E;
2360 case X86::COND_L: return X86::COND_GE;
2361 case X86::COND_LE: return X86::COND_G;
2362 case X86::COND_G: return X86::COND_LE;
2363 case X86::COND_GE: return X86::COND_L;
2364 case X86::COND_B: return X86::COND_AE;
2365 case X86::COND_BE: return X86::COND_A;
2366 case X86::COND_A: return X86::COND_BE;
2367 case X86::COND_AE: return X86::COND_B;
2368 case X86::COND_S: return X86::COND_NS;
2369 case X86::COND_NS: return X86::COND_S;
2370 case X86::COND_P: return X86::COND_NP;
2371 case X86::COND_NP: return X86::COND_P;
2372 case X86::COND_O: return X86::COND_NO;
2373 case X86::COND_NO: return X86::COND_O;
2374 }
2375}
2376
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2377/// getSwappedCondition - assume the flags are set by MI(a,b), return
2378/// the condition code if we modify the instructions such that flags are
2379/// set by MI(b,a).
Benjamin Kramerabbfe692012-07-13 13:25:15 +0000[diff] [blame]2380static X86::CondCode getSwappedCondition(X86::CondCode CC) {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2381 switch (CC) {
2382 default: return X86::COND_INVALID;
2383 case X86::COND_E: return X86::COND_E;
2384 case X86::COND_NE: return X86::COND_NE;
2385 case X86::COND_L: return X86::COND_G;
2386 case X86::COND_LE: return X86::COND_GE;
2387 case X86::COND_G: return X86::COND_L;
2388 case X86::COND_GE: return X86::COND_LE;
2389 case X86::COND_B: return X86::COND_A;
2390 case X86::COND_BE: return X86::COND_AE;
2391 case X86::COND_A: return X86::COND_B;
2392 case X86::COND_AE: return X86::COND_BE;
2393 }
2394}
2395
2396/// getSETFromCond - Return a set opcode for the given condition and
2397/// whether it has memory operand.
2398static unsigned getSETFromCond(X86::CondCode CC,
2399 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2400 static const uint16_t Opc[16][2] = {
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2401 { X86::SETAr, X86::SETAm },
2402 { X86::SETAEr, X86::SETAEm },
2403 { X86::SETBr, X86::SETBm },
2404 { X86::SETBEr, X86::SETBEm },
2405 { X86::SETEr, X86::SETEm },
2406 { X86::SETGr, X86::SETGm },
2407 { X86::SETGEr, X86::SETGEm },
2408 { X86::SETLr, X86::SETLm },
2409 { X86::SETLEr, X86::SETLEm },
2410 { X86::SETNEr, X86::SETNEm },
2411 { X86::SETNOr, X86::SETNOm },
2412 { X86::SETNPr, X86::SETNPm },
2413 { X86::SETNSr, X86::SETNSm },
2414 { X86::SETOr, X86::SETOm },
2415 { X86::SETPr, X86::SETPm },
2416 { X86::SETSr, X86::SETSm }
2417 };
2418
2419 assert(CC < 16 && "Can only handle standard cond codes");
2420 return Opc[CC][HasMemoryOperand ? 1 : 0];
2421}
2422
2423/// getCMovFromCond - Return a cmov opcode for the given condition,
2424/// register size in bytes, and operand type.
2425static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2426 bool HasMemoryOperand) {
Craig Topperbfcfdeb2012-08-21 08:23:21 +0000[diff] [blame]2427 static const uint16_t Opc[32][3] = {
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2428 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2429 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2430 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2431 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2432 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2433 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2434 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2435 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2436 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2437 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2438 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2439 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2440 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2441 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2442 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2443 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2444 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2445 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2446 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2447 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2448 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2449 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2450 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2451 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2452 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2453 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2454 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2455 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2456 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2457 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2458 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2459 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2460 };
2461
2462 assert(CC < 16 && "Can only handle standard cond codes");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2463 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2464 switch(RegBytes) {
2465 default: llvm_unreachable("Illegal register size!");
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2466 case 2: return Opc[Idx][0];
2467 case 4: return Opc[Idx][1];
2468 case 8: return Opc[Idx][2];
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2469 }
2470}
2471
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2472bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2473 if (!MI->isTerminator()) return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2474
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2475 // Conditional branch is a special case.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2476 if (MI->isBranch() && !MI->isBarrier())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2477 return true;
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2478 if (!MI->isPredicable())
Chris Lattnera98c6792008-01-07 01:56:04 +0000[diff] [blame]2479 return true;
2480 return !isPredicated(MI);
Dale Johannesen616627b2007-06-14 22:03:45 +0000[diff] [blame]2481}
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]2482
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2483bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2484 MachineBasicBlock *&TBB,
2485 MachineBasicBlock *&FBB,
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2486 SmallVectorImpl<MachineOperand> &Cond,
2487 bool AllowModify) const {
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2488 // Start from the bottom of the block and work up, examining the
2489 // terminator instructions.
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2490 MachineBasicBlock::iterator I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2491 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2492 while (I != MBB.begin()) {
2493 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2494 if (I->isDebugValue())
2495 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2496
2497 // Working from the bottom, when we see a non-terminator instruction, we're
2498 // done.
Jakob Stoklund Olesenc30b4dd2010-07-16 17:41:44 +0000[diff] [blame]2499 if (!isUnpredicatedTerminator(I))
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2500 break;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2501
2502 // A terminator that isn't a branch can't easily be handled by this
2503 // analysis.
Evan Cheng7f8e5632011-12-07 07:15:52 +0000[diff] [blame]2504 if (!I->isBranch())
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2505 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2506
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2507 // Handle unconditional branches.
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2508 if (I->getOpcode() == X86::JMP_4) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2509 UnCondBrIter = I;
2510
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2511 if (!AllowModify) {
2512 TBB = I->getOperand(0).getMBB();
Evan Cheng2fa28112009-05-08 06:34:09 +0000[diff] [blame]2513 continue;
Evan Cheng64dfcac2009-02-09 07:14:22 +0000[diff] [blame]2514 }
2515
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2516 // If the block has any instructions after a JMP, delete them.
Chris Lattnera48f44d2009-12-03 00:50:42 +0000[diff] [blame]2517 while (llvm::next(I) != MBB.end())
2518 llvm::next(I)->eraseFromParent();
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2519
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2520 Cond.clear();
2521 FBB = 0;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2522
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2523 // Delete the JMP if it's equivalent to a fall-through.
2524 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2525 TBB = 0;
2526 I->eraseFromParent();
2527 I = MBB.end();
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2528 UnCondBrIter = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2529 continue;
2530 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2531
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2532 // TBB is used to indicate the unconditional destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2533 TBB = I->getOperand(0).getMBB();
2534 continue;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2535 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2536
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2537 // Handle conditional branches.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2538 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2539 if (BranchCode == X86::COND_INVALID)
2540 return true; // Can't handle indirect branch.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2541
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2542 // Working from the bottom, handle the first conditional branch.
2543 if (Cond.empty()) {
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2544 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2545 if (AllowModify && UnCondBrIter != MBB.end() &&
2546 MBB.isLayoutSuccessor(TargetBB)) {
2547 // If we can modify the code and it ends in something like:
2548 //
2549 // jCC L1
2550 // jmp L2
2551 // L1:
2552 // ...
2553 // L2:
2554 //
2555 // Then we can change this to:
2556 //
2557 // jnCC L2
2558 // L1:
2559 // ...
2560 // L2:
2561 //
2562 // Which is a bit more efficient.
2563 // We conditionally jump to the fall-through block.
2564 BranchCode = GetOppositeBranchCondition(BranchCode);
2565 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2566 MachineBasicBlock::iterator OldInst = I;
2567
2568 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2569 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2570 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2571 .addMBB(TargetBB);
Evan Cheng4ca4bc62010-04-13 18:50:27 +0000[diff] [blame]2572
2573 OldInst->eraseFromParent();
2574 UnCondBrIter->eraseFromParent();
2575
2576 // Restart the analysis.
2577 UnCondBrIter = MBB.end();
2578 I = MBB.end();
2579 continue;
2580 }
2581
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2582 FBB = TBB;
2583 TBB = I->getOperand(0).getMBB();
2584 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2585 continue;
2586 }
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2587
2588 // Handle subsequent conditional branches. Only handle the case where all
2589 // conditional branches branch to the same destination and their condition
2590 // opcodes fit one of the special multi-branch idioms.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2591 assert(Cond.size() == 1);
2592 assert(TBB);
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2593
2594 // Only handle the case where all conditional branches branch to the same
2595 // destination.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2596 if (TBB != I->getOperand(0).getMBB())
2597 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2598
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2599 // If the conditions are the same, we can leave them alone.
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2600 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2601 if (OldBranchCode == BranchCode)
2602 continue;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2603
2604 // If they differ, see if they fit one of the known patterns. Theoretically,
2605 // we could handle more patterns here, but we shouldn't expect to see them
2606 // if instruction selection has done a reasonable job.
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2607 if ((OldBranchCode == X86::COND_NP &&
2608 BranchCode == X86::COND_E) ||
2609 (OldBranchCode == X86::COND_E &&
2610 BranchCode == X86::COND_NP))
2611 BranchCode = X86::COND_NP_OR_E;
2612 else if ((OldBranchCode == X86::COND_P &&
2613 BranchCode == X86::COND_NE) ||
2614 (OldBranchCode == X86::COND_NE &&
2615 BranchCode == X86::COND_P))
2616 BranchCode = X86::COND_NE_OR_P;
2617 else
2618 return true;
Bill Wendling277381f2009-12-14 06:51:19 +0000[diff] [blame]2619
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2620 // Update the MachineOperand.
2621 Cond[0].setImm(BranchCode);
Chris Lattner74436002006-10-30 22:27:23 +0000[diff] [blame]2622 }
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2623
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2624 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2625}
2626
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2627unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2628 MachineBasicBlock::iterator I = MBB.end();
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2629 unsigned Count = 0;
2630
2631 while (I != MBB.begin()) {
2632 --I;
Dale Johannesen4244d122010-04-02 01:38:09 +0000[diff] [blame]2633 if (I->isDebugValue())
2634 continue;
Chris Lattner2b0a7a22010-02-11 19:25:55 +0000[diff] [blame]2635 if (I->getOpcode() != X86::JMP_4 &&
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2636 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2637 break;
2638 // Remove the branch.
2639 I->eraseFromParent();
2640 I = MBB.end();
2641 ++Count;
2642 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]2643
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2644 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2645}
2646
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2647unsigned
2648X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2649 MachineBasicBlock *FBB,
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2650 const SmallVectorImpl<MachineOperand> &Cond,
2651 DebugLoc DL) const {
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2652 // Shouldn't be a fall through.
2653 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
Chris Lattner6fca75e2006-10-21 05:34:23 +0000[diff] [blame]2654 assert((Cond.size() == 1 || Cond.size() == 0) &&
2655 "X86 branch conditions have one component!");
2656
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2657 if (Cond.empty()) {
2658 // Unconditional branch?
2659 assert(!FBB && "Unconditional branch with multiple successors!");
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2660 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
Evan Chenge20dd922007-05-18 00:18:17 +0000[diff] [blame]2661 return 1;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2662 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2663
2664 // Conditional branch.
2665 unsigned Count = 0;
2666 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2667 switch (CC) {
2668 case X86::COND_NP_OR_E:
2669 // Synthesize NP_OR_E with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2670 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2671 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2672 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2673 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2674 break;
2675 case X86::COND_NE_OR_P:
2676 // Synthesize NE_OR_P with two branches.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2677 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2678 ++Count;
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2679 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2680 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2681 break;
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2682 default: {
2683 unsigned Opc = GetCondBranchFromCond(CC);
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2684 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2685 ++Count;
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2686 }
Bill Wendling543ce1f2010-03-05 00:33:59 +0000[diff] [blame]2687 }
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2688 if (FBB) {
2689 // Two-way Conditional branch. Insert the second branch.
Stuart Hastings0125b642010-06-17 22:43:56 +0000[diff] [blame]2690 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]2691 ++Count;
2692 }
2693 return Count;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]2694}
2695
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2696bool X86InstrInfo::
2697canInsertSelect(const MachineBasicBlock &MBB,
2698 const SmallVectorImpl<MachineOperand> &Cond,
2699 unsigned TrueReg, unsigned FalseReg,
2700 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2701 // Not all subtargets have cmov instructions.
2702 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2703 return false;
2704 if (Cond.size() != 1)
2705 return false;
2706 // We cannot do the composite conditions, at least not in SSA form.
2707 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2708 return false;
2709
2710 // Check register classes.
2711 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2712 const TargetRegisterClass *RC =
2713 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2714 if (!RC)
2715 return false;
2716
2717 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2718 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2719 X86::GR32RegClass.hasSubClassEq(RC) ||
2720 X86::GR64RegClass.hasSubClassEq(RC)) {
2721 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2722 // Bridge. Probably Ivy Bridge as well.
2723 CondCycles = 2;
2724 TrueCycles = 2;
2725 FalseCycles = 2;
2726 return true;
2727 }
2728
2729 // Can't do vectors.
2730 return false;
2731}
2732
2733void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2734 MachineBasicBlock::iterator I, DebugLoc DL,
2735 unsigned DstReg,
2736 const SmallVectorImpl<MachineOperand> &Cond,
2737 unsigned TrueReg, unsigned FalseReg) const {
2738 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2739 assert(Cond.size() == 1 && "Invalid Cond array");
2740 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]2741 MRI.getRegClass(DstReg)->getSize(),
2742 false/*HasMemoryOperand*/);
Jakob Stoklund Olesen49e4d4b2012-07-04 00:09:58 +0000[diff] [blame]2743 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2744}
2745
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2746/// isHReg - Test if the given register is a physical h register.
2747static bool isHReg(unsigned Reg) {
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2748 return X86::GR8_ABCD_HRegClass.contains(Reg);
Dan Gohman7913ea52009-04-15 00:04:23 +0000[diff] [blame]2749}
2750
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2751// Try and copy between VR128/VR64 and GR64 registers.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2752static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2753 bool HasAVX) {
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2754 // SrcReg(VR128) -> DestReg(GR64)
2755 // SrcReg(VR64) -> DestReg(GR64)
2756 // SrcReg(GR64) -> DestReg(VR128)
2757 // SrcReg(GR64) -> DestReg(VR64)
2758
2759 if (X86::GR64RegClass.contains(DestReg)) {
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2760 if (X86::VR128RegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2761 // Copy from a VR128 register to a GR64 register.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2762 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2763 if (X86::VR64RegClass.contains(SrcReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2764 // Copy from a VR64 register to a GR64 register.
2765 return X86::MOVSDto64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2766 } else if (X86::GR64RegClass.contains(SrcReg)) {
2767 // Copy from a GR64 register to a VR128 register.
2768 if (X86::VR128RegClass.contains(DestReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2769 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2770 // Copy from a GR64 register to a VR64 register.
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2771 if (X86::VR64RegClass.contains(DestReg))
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2772 return X86::MOV64toSDrr;
2773 }
2774
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2775 // SrcReg(FR32) -> DestReg(GR32)
2776 // SrcReg(GR32) -> DestReg(FR32)
2777
2778 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2779 // Copy from a FR32 register to a GR32 register.
2780 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2781
2782 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2783 // Copy from a GR32 register to a FR32 register.
2784 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
Jakob Stoklund Olesenf05864a2011-09-22 22:45:24 +0000[diff] [blame]2785
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2786 return 0;
2787}
2788
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2789void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2790 MachineBasicBlock::iterator MI, DebugLoc DL,
2791 unsigned DestReg, unsigned SrcReg,
2792 bool KillSrc) const {
2793 // First deal with the normal symmetric copies.
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2794 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2795 unsigned Opc;
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2796 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2797 Opc = X86::MOV64rr;
2798 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2799 Opc = X86::MOV32rr;
2800 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2801 Opc = X86::MOV16rr;
2802 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2803 // Copying to or from a physical H register on x86-64 requires a NOREX
2804 // move. Otherwise use a normal move.
2805 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2806 TM.getSubtarget<X86Subtarget>().is64Bit()) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2807 Opc = X86::MOV8rr_NOREX;
Jakob Stoklund Olesen464fcc02011-10-07 20:15:54 +0000[diff] [blame]2808 // Both operands must be encodable without an REX prefix.
2809 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2810 "8-bit H register can not be copied outside GR8_NOREX");
2811 } else
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2812 Opc = X86::MOV8rr;
2813 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2814 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]2815 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2816 Opc = X86::VMOVAPSYrr;
Jakob Stoklund Olesenec58a432010-07-08 22:30:35 +0000[diff] [blame]2817 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2818 Opc = X86::MMX_MOVQ64rr;
Anton Korobeynikovc0b36922010-08-27 14:43:06 +0000[diff] [blame]2819 else
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2820 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2821
2822 if (Opc) {
2823 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2824 .addReg(SrcReg, getKillRegState(KillSrc));
2825 return;
2826 }
2827
2828 // Moving EFLAGS to / from another register requires a push and a pop.
2829 if (SrcReg == X86::EFLAGS) {
2830 if (X86::GR64RegClass.contains(DestReg)) {
2831 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2832 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2833 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2834 }
2835 if (X86::GR32RegClass.contains(DestReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2836 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2837 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2838 return;
2839 }
2840 }
2841 if (DestReg == X86::EFLAGS) {
2842 if (X86::GR64RegClass.contains(SrcReg)) {
2843 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2844 .addReg(SrcReg, getKillRegState(KillSrc));
2845 BuildMI(MBB, MI, DL, get(X86::POPF64));
2846 return;
Craig Topperbab0c762012-08-21 08:29:51 +0000[diff] [blame]2847 }
2848 if (X86::GR32RegClass.contains(SrcReg)) {
Jakob Stoklund Olesen930f8082010-07-08 19:46:25 +0000[diff] [blame]2849 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2850 .addReg(SrcReg, getKillRegState(KillSrc));
2851 BuildMI(MBB, MI, DL, get(X86::POPF32));
2852 return;
2853 }
2854 }
2855
2856 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2857 << " to " << RI.getName(DestReg) << '\n');
2858 llvm_unreachable("Cannot emit physreg copy instruction");
2859}
2860
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2861static unsigned getLoadStoreRegOpcode(unsigned Reg,
2862 const TargetRegisterClass *RC,
2863 bool isStackAligned,
2864 const TargetMachine &TM,
2865 bool load) {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2866 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2867 switch (RC->getSize()) {
Rafael Espindola6635f982010-07-12 03:43:04 +0000[diff] [blame]2868 default:
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2869 llvm_unreachable("Unknown spill size");
2870 case 1:
2871 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2872 if (TM.getSubtarget<X86Subtarget>().is64Bit())
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2873 // Copying to or from a physical H register on x86-64 requires a NOREX
2874 // move. Otherwise use a normal move.
2875 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2876 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2877 return load ? X86::MOV8rm : X86::MOV8mr;
2878 case 2:
2879 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2880 return load ? X86::MOV16rm : X86::MOV16mr;
2881 case 4:
2882 if (X86::GR32RegClass.hasSubClassEq(RC))
2883 return load ? X86::MOV32rm : X86::MOV32mr;
2884 if (X86::FR32RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2885 return load ?
2886 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2887 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2888 if (X86::RFP32RegClass.hasSubClassEq(RC))
2889 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2890 llvm_unreachable("Unknown 4-byte regclass");
2891 case 8:
2892 if (X86::GR64RegClass.hasSubClassEq(RC))
2893 return load ? X86::MOV64rm : X86::MOV64mr;
2894 if (X86::FR64RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2895 return load ?
2896 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
2897 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2898 if (X86::VR64RegClass.hasSubClassEq(RC))
2899 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2900 if (X86::RFP64RegClass.hasSubClassEq(RC))
2901 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2902 llvm_unreachable("Unknown 8-byte regclass");
2903 case 10:
2904 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2905 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2906 case 16: {
Jakob Stoklund Olesen56ce3a02011-06-01 15:32:10 +0000[diff] [blame]2907 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2908 // If stack is realigned we can use aligned stores.
2909 if (isStackAligned)
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2910 return load ?
2911 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
2912 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2913 else
Bruno Cardoso Lopesdb520db2011-08-31 03:04:09 +0000[diff] [blame]2914 return load ?
2915 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
2916 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
2917 }
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]2918 case 32:
2919 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2920 // If stack is realigned we can use aligned stores.
2921 if (isStackAligned)
2922 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2923 else
2924 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2925 }
2926}
2927
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2928static unsigned getStoreRegOpcode(unsigned SrcReg,
2929 const TargetRegisterClass *RC,
2930 bool isStackAligned,
2931 TargetMachine &TM) {
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2932 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2933}
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2934
Rafael Espindolae302f832010-06-12 20:13:29 +0000[diff] [blame]2935
2936static unsigned getLoadRegOpcode(unsigned DestReg,
2937 const TargetRegisterClass *RC,
2938 bool isStackAligned,
2939 const TargetMachine &TM) {
2940 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2941}
2942
2943void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
2944 MachineBasicBlock::iterator MI,
2945 unsigned SrcReg, bool isKill, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]2946 const TargetRegisterClass *RC,
2947 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]2948 const MachineFunction &MF = *MBB.getParent();
Jakob Stoklund Olesenc3c05ed2010-07-27 04:16:58 +0000[diff] [blame]2949 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
2950 "Stack slot too small for store");
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2951 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2952 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]2953 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2954 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]2955 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2956 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]2957 .addReg(SrcReg, getKillRegState(isKill));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2958}
2959
2960void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
2961 bool isKill,
2962 SmallVectorImpl<MachineOperand> &Addr,
2963 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]2964 MachineInstr::mmo_iterator MMOBegin,
2965 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2966 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2967 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2968 bool isAligned = MMOBegin != MMOEnd &&
2969 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2970 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]2971 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]2972 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2973 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]2974 MIB.addOperand(Addr[i]);
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]2975 MIB.addReg(SrcReg, getKillRegState(isKill));
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]2976 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2977 NewMIs.push_back(MIB);
2978}
2979
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2980
2981void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]2982 MachineBasicBlock::iterator MI,
2983 unsigned DestReg, int FrameIdx,
Evan Chengefb126a2010-05-06 19:06:44 +0000[diff] [blame]2984 const TargetRegisterClass *RC,
2985 const TargetRegisterInfo *TRI) const {
Anton Korobeynikovb7a49922008-07-19 06:30:51 +0000[diff] [blame]2986 const MachineFunction &MF = *MBB.getParent();
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]2987 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2988 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Chengee9b90a2011-06-23 01:53:43 +0000[diff] [blame]2989 RI.canRealignStack(MF);
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]2990 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Dale Johannesene5a41342010-01-26 00:03:12 +0000[diff] [blame]2991 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]2992 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]2993}
2994
2995void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]2996 SmallVectorImpl<MachineOperand> &Addr,
2997 const TargetRegisterClass *RC,
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]2998 MachineInstr::mmo_iterator MMOBegin,
2999 MachineInstr::mmo_iterator MMOEnd,
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3000 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]3001 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3002 bool isAligned = MMOBegin != MMOEnd &&
3003 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman29869722009-04-27 16:41:36 +0000[diff] [blame]3004 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Chris Lattner6f306d72010-04-02 20:16:16 +0000[diff] [blame]3005 DebugLoc DL;
Dale Johannesen6b8c76a2009-02-12 23:08:38 +0000[diff] [blame]3006 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3007 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3008 MIB.addOperand(Addr[i]);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]3009 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersoneee14602008-01-01 21:11:32 +0000[diff] [blame]3010 NewMIs.push_back(MIB);
3011}
3012
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3013bool X86InstrInfo::
3014analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3015 int &CmpMask, int &CmpValue) const {
3016 switch (MI->getOpcode()) {
3017 default: break;
3018 case X86::CMP64ri32:
3019 case X86::CMP64ri8:
3020 case X86::CMP32ri:
3021 case X86::CMP32ri8:
3022 case X86::CMP16ri:
3023 case X86::CMP16ri8:
3024 case X86::CMP8ri:
3025 SrcReg = MI->getOperand(0).getReg();
3026 SrcReg2 = 0;
3027 CmpMask = ~0;
3028 CmpValue = MI->getOperand(1).getImm();
3029 return true;
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3030 // A SUB can be used to perform comparison.
3031 case X86::SUB64rm:
3032 case X86::SUB32rm:
3033 case X86::SUB16rm:
3034 case X86::SUB8rm:
3035 SrcReg = MI->getOperand(1).getReg();
3036 SrcReg2 = 0;
3037 CmpMask = ~0;
3038 CmpValue = 0;
3039 return true;
3040 case X86::SUB64rr:
3041 case X86::SUB32rr:
3042 case X86::SUB16rr:
3043 case X86::SUB8rr:
3044 SrcReg = MI->getOperand(1).getReg();
3045 SrcReg2 = MI->getOperand(2).getReg();
3046 CmpMask = ~0;
3047 CmpValue = 0;
3048 return true;
3049 case X86::SUB64ri32:
3050 case X86::SUB64ri8:
3051 case X86::SUB32ri:
3052 case X86::SUB32ri8:
3053 case X86::SUB16ri:
3054 case X86::SUB16ri8:
3055 case X86::SUB8ri:
3056 SrcReg = MI->getOperand(1).getReg();
3057 SrcReg2 = 0;
3058 CmpMask = ~0;
3059 CmpValue = MI->getOperand(2).getImm();
3060 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3061 case X86::CMP64rr:
3062 case X86::CMP32rr:
3063 case X86::CMP16rr:
3064 case X86::CMP8rr:
3065 SrcReg = MI->getOperand(0).getReg();
3066 SrcReg2 = MI->getOperand(1).getReg();
3067 CmpMask = ~0;
3068 CmpValue = 0;
3069 return true;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3070 case X86::TEST8rr:
3071 case X86::TEST16rr:
3072 case X86::TEST32rr:
3073 case X86::TEST64rr:
3074 SrcReg = MI->getOperand(0).getReg();
3075 if (MI->getOperand(1).getReg() != SrcReg) return false;
3076 // Compare against zero.
3077 SrcReg2 = 0;
3078 CmpMask = ~0;
3079 CmpValue = 0;
3080 return true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3081 }
3082 return false;
3083}
3084
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3085/// isRedundantFlagInstr - check whether the first instruction, whose only
3086/// purpose is to update flags, can be made redundant.
3087/// CMPrr can be made redundant by SUBrr if the operands are the same.
3088/// This function can be extended later on.
3089/// SrcReg, SrcRegs: register operands for FlagI.
3090/// ImmValue: immediate for FlagI if it takes an immediate.
3091inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3092 unsigned SrcReg2, int ImmValue,
3093 MachineInstr *OI) {
3094 if (((FlagI->getOpcode() == X86::CMP64rr &&
3095 OI->getOpcode() == X86::SUB64rr) ||
3096 (FlagI->getOpcode() == X86::CMP32rr &&
3097 OI->getOpcode() == X86::SUB32rr)||
3098 (FlagI->getOpcode() == X86::CMP16rr &&
3099 OI->getOpcode() == X86::SUB16rr)||
3100 (FlagI->getOpcode() == X86::CMP8rr &&
3101 OI->getOpcode() == X86::SUB8rr)) &&
3102 ((OI->getOperand(1).getReg() == SrcReg &&
3103 OI->getOperand(2).getReg() == SrcReg2) ||
3104 (OI->getOperand(1).getReg() == SrcReg2 &&
3105 OI->getOperand(2).getReg() == SrcReg)))
3106 return true;
3107
3108 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3109 OI->getOpcode() == X86::SUB64ri32) ||
3110 (FlagI->getOpcode() == X86::CMP64ri8 &&
3111 OI->getOpcode() == X86::SUB64ri8) ||
3112 (FlagI->getOpcode() == X86::CMP32ri &&
3113 OI->getOpcode() == X86::SUB32ri) ||
3114 (FlagI->getOpcode() == X86::CMP32ri8 &&
3115 OI->getOpcode() == X86::SUB32ri8) ||
3116 (FlagI->getOpcode() == X86::CMP16ri &&
3117 OI->getOpcode() == X86::SUB16ri) ||
3118 (FlagI->getOpcode() == X86::CMP16ri8 &&
3119 OI->getOpcode() == X86::SUB16ri8) ||
3120 (FlagI->getOpcode() == X86::CMP8ri &&
3121 OI->getOpcode() == X86::SUB8ri)) &&
3122 OI->getOperand(1).getReg() == SrcReg &&
3123 OI->getOperand(2).getImm() == ImmValue)
3124 return true;
3125 return false;
3126}
3127
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3128/// isDefConvertible - check whether the definition can be converted
3129/// to remove a comparison against zero.
3130inline static bool isDefConvertible(MachineInstr *MI) {
3131 switch (MI->getOpcode()) {
3132 default: return false;
3133 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3134 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3135 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3136 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3137 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3138 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3139 case X86::DEC64m: case X86::DEC32m: case X86::DEC16m: case X86::DEC8m:
3140 case X86::DEC64_32r: case X86::DEC64_16r:
3141 case X86::DEC64_32m: case X86::DEC64_16m:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3142 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3143 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3144 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3145 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3146 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3147 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3148 case X86::INC64m: case X86::INC32m: case X86::INC16m: case X86::INC8m:
3149 case X86::INC64_32r: case X86::INC64_16r:
3150 case X86::INC64_32m: case X86::INC64_16m:
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3151 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3152 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3153 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3154 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3155 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3156 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3157 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3158 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3159 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3160 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3161 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3162 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3163 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3164 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3165 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3166 return true;
3167 }
3168}
3169
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3170/// optimizeCompareInstr - Check if there exists an earlier instruction that
3171/// operates on the same source operands and sets flags in the same way as
3172/// Compare; remove Compare if possible.
3173bool X86InstrInfo::
3174optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3175 int CmpMask, int CmpValue,
3176 const MachineRegisterInfo *MRI) const {
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3177 // Check whether we can replace SUB with CMP.
3178 unsigned NewOpcode = 0;
3179 switch (CmpInstr->getOpcode()) {
3180 default: break;
3181 case X86::SUB64ri32:
3182 case X86::SUB64ri8:
3183 case X86::SUB32ri:
3184 case X86::SUB32ri8:
3185 case X86::SUB16ri:
3186 case X86::SUB16ri8:
3187 case X86::SUB8ri:
3188 case X86::SUB64rm:
3189 case X86::SUB32rm:
3190 case X86::SUB16rm:
3191 case X86::SUB8rm:
3192 case X86::SUB64rr:
3193 case X86::SUB32rr:
3194 case X86::SUB16rr:
3195 case X86::SUB8rr: {
3196 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3197 return false;
3198 // There is no use of the destination register, we can replace SUB with CMP.
3199 switch (CmpInstr->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]3200 default: llvm_unreachable("Unreachable!");
Manman Ren1be131b2012-08-08 00:51:41 +0000[diff] [blame]3201 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3202 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3203 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3204 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3205 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3206 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3207 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3208 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3209 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3210 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3211 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3212 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3213 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3214 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3215 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3216 }
3217 CmpInstr->setDesc(get(NewOpcode));
3218 CmpInstr->RemoveOperand(0);
3219 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3220 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3221 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3222 return false;
3223 }
3224 }
3225
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3226 // Get the unique definition of SrcReg.
3227 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3228 if (!MI) return false;
3229
3230 // CmpInstr is the first instruction of the BB.
3231 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3232
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3233 // If we are comparing against zero, check whether we can use MI to update
3234 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3235 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3236 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3237 !isDefConvertible(MI)))
3238 return false;
3239
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3240 // We are searching for an earlier instruction that can make CmpInstr
3241 // redundant and that instruction will be saved in Sub.
3242 MachineInstr *Sub = NULL;
3243 const TargetRegisterInfo *TRI = &getRegisterInfo();
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3244
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3245 // We iterate backward, starting from the instruction before CmpInstr and
3246 // stop when reaching the definition of a source register or done with the BB.
3247 // RI points to the instruction before CmpInstr.
3248 // If the definition is in this basic block, RE points to the definition;
3249 // otherwise, RE is the rend of the basic block.
3250 MachineBasicBlock::reverse_iterator
3251 RI = MachineBasicBlock::reverse_iterator(I),
3252 RE = CmpInstr->getParent() == MI->getParent() ?
3253 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3254 CmpInstr->getParent()->rend();
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3255 MachineInstr *Movr0Inst = 0;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3256 for (; RI != RE; ++RI) {
3257 MachineInstr *Instr = &*RI;
3258 // Check whether CmpInstr can be made redundant by the current instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3259 if (!IsCmpZero &&
3260 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3261 Sub = Instr;
3262 break;
3263 }
3264
3265 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3266 Instr->readsRegister(X86::EFLAGS, TRI)) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3267 // This instruction modifies or uses EFLAGS.
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3268
3269 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3270 // They are safe to move up, if the definition to EFLAGS is dead and
3271 // earlier instructions do not read or write EFLAGS.
3272 if (!Movr0Inst && (Instr->getOpcode() == X86::MOV8r0 ||
3273 Instr->getOpcode() == X86::MOV16r0 ||
3274 Instr->getOpcode() == X86::MOV32r0 ||
3275 Instr->getOpcode() == X86::MOV64r0) &&
3276 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3277 Movr0Inst = Instr;
3278 continue;
3279 }
3280
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3281 // We can't remove CmpInstr.
3282 return false;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3283 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3284 }
3285
3286 // Return false if no candidates exist.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3287 if (!IsCmpZero && !Sub)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3288 return false;
3289
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3290 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3291 Sub->getOperand(2).getReg() == SrcReg);
3292
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3293 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3294 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3295 // If we are done with the basic block, we need to check whether EFLAGS is
3296 // live-out.
3297 bool IsSafe = false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3298 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3299 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3300 for (++I; I != E; ++I) {
3301 const MachineInstr &Instr = *I;
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3302 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3303 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3304 // We should check the usage if this instruction uses and updates EFLAGS.
3305 if (!UseEFLAGS && ModifyEFLAGS) {
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3306 // It is safe to remove CmpInstr if EFLAGS is updated again.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3307 IsSafe = true;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3308 break;
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3309 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3310 if (!UseEFLAGS && !ModifyEFLAGS)
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3311 continue;
3312
3313 // EFLAGS is used by this instruction.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3314 X86::CondCode OldCC;
3315 bool OpcIsSET = false;
3316 if (IsCmpZero || IsSwapped) {
3317 // We decode the condition code from opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3318 if (Instr.isBranch())
3319 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3320 else {
3321 OldCC = getCondFromSETOpc(Instr.getOpcode());
3322 if (OldCC != X86::COND_INVALID)
3323 OpcIsSET = true;
3324 else
Michael Liao32376622012-09-20 03:06:15 +0000[diff] [blame]3325 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3326 }
3327 if (OldCC == X86::COND_INVALID) return false;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3328 }
3329 if (IsCmpZero) {
3330 switch (OldCC) {
3331 default: break;
3332 case X86::COND_A: case X86::COND_AE:
3333 case X86::COND_B: case X86::COND_BE:
3334 case X86::COND_G: case X86::COND_GE:
3335 case X86::COND_L: case X86::COND_LE:
3336 case X86::COND_O: case X86::COND_NO:
3337 // CF and OF are used, we can't perform this optimization.
3338 return false;
3339 }
3340 } else if (IsSwapped) {
3341 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3342 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3343 // We swap the condition code and synthesize the new opcode.
Manman Ren5f6fa422012-07-09 18:57:12 +0000[diff] [blame]3344 X86::CondCode NewCC = getSwappedCondition(OldCC);
3345 if (NewCC == X86::COND_INVALID) return false;
3346
3347 // Synthesize the new opcode.
3348 bool HasMemoryOperand = Instr.hasOneMemOperand();
3349 unsigned NewOpc;
3350 if (Instr.isBranch())
3351 NewOpc = GetCondBranchFromCond(NewCC);
3352 else if(OpcIsSET)
3353 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3354 else {
3355 unsigned DstReg = Instr.getOperand(0).getReg();
3356 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3357 HasMemoryOperand);
3358 }
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3359
3360 // Push the MachineInstr to OpsToUpdate.
3361 // If it is safe to remove CmpInstr, the condition code of these
3362 // instructions will be modified.
3363 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3364 }
Manman Ren32367c02012-07-28 03:15:46 +0000[diff] [blame]3365 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3366 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3367 IsSafe = true;
3368 break;
3369 }
3370 }
3371
3372 // If EFLAGS is not killed nor re-defined, we should check whether it is
3373 // live-out. If it is live-out, do not optimize.
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3374 if ((IsCmpZero || IsSwapped) && !IsSafe) {
Manman Renbb360742012-07-07 03:34:46 +0000[diff] [blame]3375 MachineBasicBlock *MBB = CmpInstr->getParent();
3376 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3377 SE = MBB->succ_end(); SI != SE; ++SI)
3378 if ((*SI)->isLiveIn(X86::EFLAGS))
3379 return false;
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3380 }
3381
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3382 // The instruction to be updated is either Sub or MI.
3383 Sub = IsCmpZero ? MI : Sub;
Manman Ren1553ce02012-07-11 19:35:12 +0000[diff] [blame]3384 // Move Movr0Inst to the place right before Sub.
3385 if (Movr0Inst) {
3386 Sub->getParent()->remove(Movr0Inst);
3387 Sub->getParent()->insert(MachineBasicBlock::iterator(Sub), Movr0Inst);
3388 }
3389
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3390 // Make sure Sub instruction defines EFLAGS and mark the def live.
3391 unsigned LastOperand = Sub->getNumOperands() - 1;
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3392 assert(Sub->getNumOperands() >= 2 &&
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3393 Sub->getOperand(LastOperand).isReg() &&
3394 Sub->getOperand(LastOperand).getReg() == X86::EFLAGS &&
Manman Rend0a4ee82012-07-18 21:40:01 +0000[diff] [blame]3395 "EFLAGS should be the last operand of SUB, ADD, OR, XOR, AND");
Jan Wen Voung4ce1d7b2012-09-17 22:04:23 +0000[diff] [blame]3396 Sub->getOperand(LastOperand).setIsDef(true);
3397 Sub->getOperand(LastOperand).setIsDead(false);
Manman Renc9656732012-07-06 17:36:20 +0000[diff] [blame]3398 CmpInstr->eraseFromParent();
3399
3400 // Modify the condition code of instructions in OpsToUpdate.
3401 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3402 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3403 return true;
3404}
3405
Manman Ren5759d012012-08-02 00:56:42 +0000[diff] [blame]3406/// optimizeLoadInstr - Try to remove the load by folding it to a register
3407/// operand at the use. We fold the load instructions if load defines a virtual
3408/// register, the virtual register is used once in the same BB, and the
3409/// instructions in-between do not load or store, and have no side effects.
3410MachineInstr* X86InstrInfo::
3411optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3412 unsigned &FoldAsLoadDefReg,
3413 MachineInstr *&DefMI) const {
3414 if (FoldAsLoadDefReg == 0)
3415 return 0;
3416 // To be conservative, if there exists another load, clear the load candidate.
3417 if (MI->mayLoad()) {
3418 FoldAsLoadDefReg = 0;
3419 return 0;
3420 }
3421
3422 // Check whether we can move DefMI here.
3423 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3424 assert(DefMI);
3425 bool SawStore = false;
3426 if (!DefMI->isSafeToMove(this, 0, SawStore))
3427 return 0;
3428
3429 // We try to commute MI if possible.
3430 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3431 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3432 // Collect information about virtual register operands of MI.
3433 unsigned SrcOperandId = 0;
3434 bool FoundSrcOperand = false;
3435 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3436 MachineOperand &MO = MI->getOperand(i);
3437 if (!MO.isReg())
3438 continue;
3439 unsigned Reg = MO.getReg();
3440 if (Reg != FoldAsLoadDefReg)
3441 continue;
3442 // Do not fold if we have a subreg use or a def or multiple uses.
3443 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3444 return 0;
3445
3446 SrcOperandId = i;
3447 FoundSrcOperand = true;
3448 }
3449 if (!FoundSrcOperand) return 0;
3450
3451 // Check whether we can fold the def into SrcOperandId.
3452 SmallVector<unsigned, 8> Ops;
3453 Ops.push_back(SrcOperandId);
3454 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3455 if (FoldMI) {
3456 FoldAsLoadDefReg = 0;
3457 return FoldMI;
3458 }
3459
3460 if (Idx == 1) {
3461 // MI was changed but it didn't help, commute it back!
3462 commuteInstruction(MI, false);
3463 return 0;
3464 }
3465
3466 // Check whether we can commute MI and enable folding.
3467 if (MI->isCommutable()) {
3468 MachineInstr *NewMI = commuteInstruction(MI, false);
3469 // Unable to commute.
3470 if (!NewMI) return 0;
3471 if (NewMI != MI) {
3472 // New instruction. It doesn't need to be kept.
3473 NewMI->eraseFromParent();
3474 return 0;
3475 }
3476 }
3477 }
3478 return 0;
3479}
3480
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3481/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3482/// instruction with two undef reads of the register being defined. This is
3483/// used for mapping:
3484/// %xmm4 = V_SET0
3485/// to:
3486/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3487///
3488static bool Expand2AddrUndef(MachineInstr *MI, const MCInstrDesc &Desc) {
3489 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3490 unsigned Reg = MI->getOperand(0).getReg();
3491 MI->setDesc(Desc);
3492
3493 // MachineInstr::addOperand() will insert explicit operands before any
3494 // implicit operands.
3495 MachineInstrBuilder(MI).addReg(Reg, RegState::Undef)
3496 .addReg(Reg, RegState::Undef);
3497 // But we don't trust that.
3498 assert(MI->getOperand(1).getReg() == Reg &&
3499 MI->getOperand(2).getReg() == Reg && "Misplaced operand");
3500 return true;
3501}
3502
3503bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3504 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3505 switch (MI->getOpcode()) {
3506 case X86::V_SET0:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3507 case X86::FsFLD0SS:
3508 case X86::FsFLD0SD:
Jakob Stoklund Olesen024130892011-11-07 19:15:58 +0000[diff] [blame]3509 return Expand2AddrUndef(MI, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]3510 case X86::AVX_SET0:
3511 assert(HasAVX && "AVX not supported");
3512 return Expand2AddrUndef(MI, get(X86::VXORPSYrr));
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3513 case X86::V_SETALLONES:
3514 return Expand2AddrUndef(MI, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
3515 case X86::AVX2_SETALLONES:
3516 return Expand2AddrUndef(MI, get(X86::VPCMPEQDYrr));
Jakob Stoklund Olesen729abd32011-10-08 18:28:28 +0000[diff] [blame]3517 case X86::TEST8ri_NOREX:
3518 MI->setDesc(get(X86::TEST8ri));
3519 return true;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3520 }
3521 return false;
3522}
3523
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3524MachineInstr*
3525X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
Evan Cheng250e9172010-04-29 01:13:30 +0000[diff] [blame]3526 int FrameIx, uint64_t Offset,
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3527 const MDNode *MDPtr,
3528 DebugLoc DL) const {
Evan Chenged69b382010-04-26 07:38:55 +0000[diff] [blame]3529 X86AddressMode AM;
3530 AM.BaseType = X86AddressMode::FrameIndexBase;
3531 AM.Base.FrameIndex = FrameIx;
3532 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
3533 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
3534 return &*MIB;
3535}
3536
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3537static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3538 const SmallVectorImpl<MachineOperand> &MOs,
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3539 MachineInstr *MI,
3540 const TargetInstrInfo &TII) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3541 // Create the base instruction with the memory operand as the first part.
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3542 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3543 MI->getDebugLoc(), true);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3544 MachineInstrBuilder MIB(NewMI);
3545 unsigned NumAddrOps = MOs.size();
3546 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3547 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3548 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3549 addOffset(MIB, 0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3550
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3551 // Loop over the rest of the ri operands, converting them over.
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3552 unsigned NumOps = MI->getDesc().getNumOperands()-2;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3553 for (unsigned i = 0; i != NumOps; ++i) {
3554 MachineOperand &MO = MI->getOperand(i+2);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3555 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3556 }
3557 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3558 MachineOperand &MO = MI->getOperand(i);
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3559 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3560 }
3561 return MIB;
3562}
3563
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3564static MachineInstr *FuseInst(MachineFunction &MF,
3565 unsigned Opcode, unsigned OpNo,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3566 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3567 MachineInstr *MI, const TargetInstrInfo &TII) {
Bill Wendlinge3c78362009-02-03 00:55:04 +0000[diff] [blame]3568 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3569 MI->getDebugLoc(), true);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3570 MachineInstrBuilder MIB(NewMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3571
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3572 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3573 MachineOperand &MO = MI->getOperand(i);
3574 if (i == OpNo) {
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3575 assert(MO.isReg() && "Expected to fold into reg operand!");
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3576 unsigned NumAddrOps = MOs.size();
3577 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3578 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3579 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3580 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3581 } else {
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3582 MIB.addOperand(MO);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3583 }
3584 }
3585 return MIB;
3586}
3587
3588static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
Dan Gohman906152a2009-01-05 17:59:02 +0000[diff] [blame]3589 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3590 MachineInstr *MI) {
Dan Gohman3b460302008-07-07 23:14:23 +0000[diff] [blame]3591 MachineFunction &MF = *MI->getParent()->getParent();
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]3592 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3593
3594 unsigned NumAddrOps = MOs.size();
3595 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]3596 MIB.addOperand(MOs[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3597 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3598 addOffset(MIB, 0);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3599 return MIB.addImm(0);
3600}
3601
3602MachineInstr*
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3603X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3604 MachineInstr *MI, unsigned i,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3605 const SmallVectorImpl<MachineOperand> &MOs,
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3606 unsigned Size, unsigned Align) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3607 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3608 bool isTwoAddrFold = false;
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3609 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3610 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]3611 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3612
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]3613 // FIXME: AsmPrinter doesn't know how to handle
3614 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3615 if (MI->getOpcode() == X86::ADD32ri &&
3616 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3617 return NULL;
3618
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3619 MachineInstr *NewMI = NULL;
3620 // Folding a memory location into the two-address part of a two-address
3621 // instruction is different than folding it other places. It requires
3622 // replacing the *two* registers with the memory location.
3623 if (isTwoAddr && NumOps >= 2 && i < 2 &&
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]3624 MI->getOperand(0).isReg() &&
3625 MI->getOperand(1).isReg() &&
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3626 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3627 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3628 isTwoAddrFold = true;
3629 } else if (i == 0) { // If operand 0
Craig Topperf9115972012-08-23 04:57:36 +0000[diff] [blame]3630 unsigned Opc = 0;
3631 switch (MI->getOpcode()) {
3632 default: break;
3633 case X86::MOV64r0: Opc = X86::MOV64mi32; break;
3634 case X86::MOV32r0: Opc = X86::MOV32mi; break;
3635 case X86::MOV16r0: Opc = X86::MOV16mi; break;
3636 case X86::MOV8r0: Opc = X86::MOV8mi; break;
3637 }
3638 if (Opc)
3639 NewMI = MakeM0Inst(*this, Opc, MOs, MI);
Evan Cheng7d98a482008-07-03 09:09:37 +0000[diff] [blame]3640 if (NewMI)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3641 return NewMI;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3642
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3643 OpcodeTablePtr = &RegOp2MemOpTable0;
3644 } else if (i == 1) {
3645 OpcodeTablePtr = &RegOp2MemOpTable1;
3646 } else if (i == 2) {
3647 OpcodeTablePtr = &RegOp2MemOpTable2;
Elena Demikhovsky3cb3b002012-08-01 12:06:00 +0000[diff] [blame]3648 } else if (i == 3) {
3649 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3650 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3651
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3652 // If table selected...
3653 if (OpcodeTablePtr) {
3654 // Find the Opcode to fuse
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]3655 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3656 OpcodeTablePtr->find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3657 if (I != OpcodeTablePtr->end()) {
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3658 unsigned Opcode = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]3659 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3660 if (Align < MinAlign)
3661 return NULL;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3662 bool NarrowToMOV32rm = false;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3663 if (Size) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]3664 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3665 if (Size < RCSize) {
3666 // Check if it's safe to fold the load. If the size of the object is
3667 // narrower than the load width, then it's not.
3668 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3669 return NULL;
3670 // If this is a 64-bit load, but the spill slot is 32, then we can do
3671 // a 32-bit load which is implicitly zero-extended. This likely is due
3672 // to liveintervalanalysis remat'ing a load from stack slot.
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3673 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3674 return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3675 Opcode = X86::MOV32rm;
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3676 NarrowToMOV32rm = true;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3677 }
3678 }
3679
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3680 if (isTwoAddrFold)
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3681 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3682 else
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3683 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3684
3685 if (NarrowToMOV32rm) {
3686 // If this is the special case where we use a MOV32rm to load a 32-bit
3687 // value and zero-extend the top bits. Change the destination register
3688 // to a 32-bit one.
3689 unsigned DstReg = NewMI->getOperand(0).getReg();
3690 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3691 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3692 X86::sub_32bit));
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3693 else
Jakob Stoklund Olesen9340ea52010-05-24 14:48:17 +0000[diff] [blame]3694 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
Evan Cheng74a32312009-09-11 01:01:31 +0000[diff] [blame]3695 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3696 return NewMI;
3697 }
3698 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3699
3700 // No fusion
Jakob Stoklund Olesen51702ec2010-07-09 20:43:09 +0000[diff] [blame]3701 if (PrintFailedFusing && !MI->isCopy())
David Greened589daf2010-01-05 01:29:29 +0000[diff] [blame]3702 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3703 return NULL;
3704}
3705
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3706/// hasPartialRegUpdate - Return true for all instructions that only update
3707/// the first 32 or 64-bits of the destination register and leave the rest
3708/// unmodified. This can be used to avoid folding loads if the instructions
3709/// only update part of the destination register, and the non-updated part is
3710/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3711/// instructions breaks the partial register dependency and it can improve
3712/// performance. e.g.:
3713///
3714/// movss (%rdi), %xmm0
3715/// cvtss2sd %xmm0, %xmm0
3716///
3717/// Instead of
3718/// cvtss2sd (%rdi), %xmm0
3719///
Bruno Cardoso Lopes7b435682011-09-15 23:04:24 +0000[diff] [blame]3720/// FIXME: This should be turned into a TSFlags.
3721///
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3722static bool hasPartialRegUpdate(unsigned Opcode) {
3723 switch (Opcode) {
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]3724 case X86::CVTSI2SSrr:
3725 case X86::CVTSI2SS64rr:
3726 case X86::CVTSI2SDrr:
3727 case X86::CVTSI2SD64rr:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3728 case X86::CVTSD2SSrr:
3729 case X86::Int_CVTSD2SSrr:
3730 case X86::CVTSS2SDrr:
3731 case X86::Int_CVTSS2SDrr:
3732 case X86::RCPSSr:
3733 case X86::RCPSSr_Int:
3734 case X86::ROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3735 case X86::ROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3736 case X86::ROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3737 case X86::ROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3738 case X86::RSQRTSSr:
3739 case X86::RSQRTSSr_Int:
3740 case X86::SQRTSSr:
3741 case X86::SQRTSSr_Int:
3742 // AVX encoded versions
3743 case X86::VCVTSD2SSrr:
3744 case X86::Int_VCVTSD2SSrr:
3745 case X86::VCVTSS2SDrr:
3746 case X86::Int_VCVTSS2SDrr:
3747 case X86::VRCPSSr:
3748 case X86::VROUNDSDr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3749 case X86::VROUNDSDr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3750 case X86::VROUNDSSr:
Benjamin Kramer2dc5dec2011-12-09 15:43:55 +0000[diff] [blame]3751 case X86::VROUNDSSr_Int:
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3752 case X86::VRSQRTSSr:
3753 case X86::VSQRTSSr:
3754 return true;
3755 }
3756
3757 return false;
3758}
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3759
Jakob Stoklund Olesenf8ad3362011-11-15 01:15:30 +0000[diff] [blame]3760/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3761/// instructions we would like before a partial register update.
3762unsigned X86InstrInfo::
3763getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3764 const TargetRegisterInfo *TRI) const {
3765 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3766 return 0;
3767
3768 // If MI is marked as reading Reg, the partial register update is wanted.
3769 const MachineOperand &MO = MI->getOperand(0);
3770 unsigned Reg = MO.getReg();
3771 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3772 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3773 return 0;
3774 } else {
3775 if (MI->readsRegister(Reg, TRI))
3776 return 0;
3777 }
3778
3779 // If any of the preceding 16 instructions are reading Reg, insert a
3780 // dependency breaking instruction. The magic number is based on a few
3781 // Nehalem experiments.
3782 return 16;
3783}
3784
3785void X86InstrInfo::
3786breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3787 const TargetRegisterInfo *TRI) const {
3788 unsigned Reg = MI->getOperand(OpNum).getReg();
3789 if (X86::VR128RegClass.contains(Reg)) {
3790 // These instructions are all floating point domain, so xorps is the best
3791 // choice.
3792 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3793 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3794 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3795 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3796 } else if (X86::VR256RegClass.contains(Reg)) {
3797 // Use vxorps to clear the full ymm register.
3798 // It wants to read and write the xmm sub-register.
3799 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3800 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3801 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3802 .addReg(Reg, RegState::ImplicitDefine);
3803 } else
3804 return;
3805 MI->addRegisterKilled(Reg, TRI, true);
3806}
3807
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3808MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3809 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3810 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3811 int FrameIndex) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3812 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3813 if (NoFusing) return NULL;
3814
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3815 // Unless optimizing for size, don't fold to avoid partial
3816 // register update stalls
3817 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3818 hasPartialRegUpdate(MI->getOpcode()))
3819 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]3820
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3821 const MachineFrameInfo *MFI = MF.getFrameInfo();
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3822 unsigned Size = MFI->getObjectSize(FrameIndex);
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3823 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3824 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3825 unsigned NewOpc = 0;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3826 unsigned RCSize = 0;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3827 switch (MI->getOpcode()) {
3828 default: return NULL;
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3829 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
Dan Gohman887dd1c2010-05-18 21:42:03 +0000[diff] [blame]3830 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
3831 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
3832 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3833 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3834 // Check if it's safe to fold the load. If the size of the object is
3835 // narrower than the load width, then it's not.
3836 if (Size < RCSize)
3837 return NULL;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3838 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]3839 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3840 MI->getOperand(1).ChangeToImmediate(0);
3841 } else if (Ops.size() != 1)
3842 return NULL;
3843
3844 SmallVector<MachineOperand,4> MOs;
3845 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3846 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3847}
3848
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3849MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3850 MachineInstr *MI,
Evan Cheng9e0c7f22009-07-15 06:10:07 +0000[diff] [blame]3851 const SmallVectorImpl<unsigned> &Ops,
Dan Gohman3f86b512008-12-03 18:43:12 +0000[diff] [blame]3852 MachineInstr *LoadMI) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3853 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3854 if (NoFusing) return NULL;
3855
Bruno Cardoso Lopes6b302952011-09-15 21:42:23 +0000[diff] [blame]3856 // Unless optimizing for size, don't fold to avoid partial
3857 // register update stalls
3858 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3859 hasPartialRegUpdate(MI->getOpcode()))
3860 return 0;
Evan Cheng4cf30b72009-12-18 07:40:29 +0000[diff] [blame]3861
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]3862 // Determine the alignment of the load.
Evan Cheng3b3286d2008-02-08 21:20:40 +0000[diff] [blame]3863 unsigned Alignment = 0;
Dan Gohman9a542a42008-07-12 00:10:52 +0000[diff] [blame]3864 if (LoadMI->hasOneMemOperand())
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]3865 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3866 else
3867 switch (LoadMI->getOpcode()) {
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]3868 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]3869 case X86::AVX_SET0:
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]3870 Alignment = 32;
3871 break;
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3872 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3873 case X86::V_SETALLONES:
3874 Alignment = 16;
3875 break;
3876 case X86::FsFLD0SD:
3877 Alignment = 8;
3878 break;
3879 case X86::FsFLD0SS:
3880 Alignment = 4;
3881 break;
3882 default:
Eli Friedman87ef3872011-06-10 01:13:01 +0000[diff] [blame]3883 return 0;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3884 }
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3885 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3886 unsigned NewOpc = 0;
3887 switch (MI->getOpcode()) {
3888 default: return NULL;
3889 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]3890 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
3891 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
3892 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3893 }
3894 // Change to CMPXXri r, 0 first.
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]3895 MI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3896 MI->getOperand(1).ChangeToImmediate(0);
3897 } else if (Ops.size() != 1)
3898 return NULL;
3899
Jakob Stoklund Olesen9c473e42010-08-11 23:08:22 +0000[diff] [blame]3900 // Make sure the subregisters match.
3901 // Otherwise we risk changing the size of the load.
3902 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
3903 return NULL;
3904
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]3905 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3906 switch (LoadMI->getOpcode()) {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3907 case X86::V_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3908 case X86::V_SETALLONES:
Craig Toppera3a65832011-11-19 22:34:59 +0000[diff] [blame]3909 case X86::AVX2_SETALLONES:
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]3910 case X86::AVX_SET0:
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3911 case X86::FsFLD0SD:
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3912 case X86::FsFLD0SS: {
Jakob Stoklund Olesendd1904e2011-09-29 05:10:54 +0000[diff] [blame]3913 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3914 // Create a constant-pool entry and operands to load from it.
3915
Dan Gohman772952f2010-03-09 03:01:40 +0000[diff] [blame]3916 // Medium and large mode can't fold loads this way.
3917 if (TM.getCodeModel() != CodeModel::Small &&
3918 TM.getCodeModel() != CodeModel::Kernel)
3919 return NULL;
3920
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3921 // x86-32 PIC requires a PIC base register for constant pools.
3922 unsigned PICBase = 0;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]3923 if (TM.getRelocationModel() == Reloc::PIC_) {
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]3924 if (TM.getSubtarget<X86Subtarget>().is64Bit())
3925 PICBase = X86::RIP;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]3926 else
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]3927 // FIXME: PICBase = getGlobalBaseReg(&MF);
Evan Chengfdd0eb42009-07-16 18:44:05 +0000[diff] [blame]3928 // This doesn't work for several reasons.
3929 // 1. GlobalBaseReg may have been spilled.
3930 // 2. It may not be live at MI.
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3931 return NULL;
Jakob Stoklund Olesenc7895d32009-07-16 21:24:13 +0000[diff] [blame]3932 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3933
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3934 // Create a constant-pool entry.
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3935 MachineConstantPool &MCP = *MF.getConstantPool();
Chris Lattner229907c2011-07-18 04:54:35 +0000[diff] [blame]3936 Type *Ty;
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]3937 unsigned Opc = LoadMI->getOpcode();
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3938 if (Opc == X86::FsFLD0SS)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3939 Ty = Type::getFloatTy(MF.getFunction()->getContext());
Jakob Stoklund Olesenbde32d32011-11-29 22:27:25 +0000[diff] [blame]3940 else if (Opc == X86::FsFLD0SD)
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3941 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
Craig Topperbd509ee2012-08-28 07:05:28 +0000[diff] [blame]3942 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
Craig Toppera4c5a472012-01-13 06:12:41 +0000[diff] [blame]3943 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3944 else
3945 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]3946
Craig Topper72f51c32012-08-28 07:30:47 +0000[diff] [blame]3947 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
Bruno Cardoso Lopes9212bf22011-07-25 23:05:32 +0000[diff] [blame]3948 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
3949 Constant::getNullValue(Ty);
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3950 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3951
3952 // Create operands to load from the constant pool entry.
3953 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
3954 MOs.push_back(MachineOperand::CreateImm(1));
3955 MOs.push_back(MachineOperand::CreateReg(0, false));
3956 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
Rafael Espindola3b2df102009-04-08 21:14:34 +0000[diff] [blame]3957 MOs.push_back(MachineOperand::CreateReg(0, false));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3958 break;
3959 }
3960 default: {
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3961 // Folding a normal load. Just copy the load's address operands.
3962 unsigned NumOps = LoadMI->getDesc().getNumOperands();
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]3963 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3964 MOs.push_back(LoadMI->getOperand(i));
Dan Gohman69499b132009-09-21 18:30:38 +0000[diff] [blame]3965 break;
3966 }
Dan Gohmancc78cdf2008-12-03 05:21:24 +0000[diff] [blame]3967 }
Evan Cheng3cad6282009-09-11 00:39:26 +0000[diff] [blame]3968 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3969}
3970
3971
Dan Gohman33332bc2008-10-16 01:49:15 +0000[diff] [blame]3972bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
3973 const SmallVectorImpl<unsigned> &Ops) const {
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3974 // Check switch flag
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3975 if (NoFusing) return 0;
3976
3977 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3978 switch (MI->getOpcode()) {
3979 default: return false;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]3980 case X86::TEST8rr:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3981 case X86::TEST16rr:
3982 case X86::TEST32rr:
3983 case X86::TEST64rr:
3984 return true;
Jakob Stoklund Olesen2348cdd2011-04-30 23:00:05 +0000[diff] [blame]3985 case X86::ADD32ri:
3986 // FIXME: AsmPrinter doesn't know how to handle
3987 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3988 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3989 return false;
3990 break;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]3991 }
3992 }
3993
3994 if (Ops.size() != 1)
3995 return false;
3996
3997 unsigned OpNum = Ops[0];
3998 unsigned Opc = MI->getOpcode();
Chris Lattner03ad8852008-01-07 07:27:27 +0000[diff] [blame]3999 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4000 bool isTwoAddr = NumOps > 1 &&
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4001 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4002
4003 // Folding a memory location into the two-address part of a two-address
4004 // instruction is different than folding it other places. It requires
4005 // replacing the *two* registers with the memory location.
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4006 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4007 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4008 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4009 } else if (OpNum == 0) { // If operand 0
4010 switch (Opc) {
Chris Lattner79c136d2009-07-14 20:19:57 +0000[diff] [blame]4011 case X86::MOV8r0:
Dan Gohmanc1195802010-01-12 04:42:54 +0000[diff] [blame]4012 case X86::MOV16r0:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4013 case X86::MOV32r0:
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4014 case X86::MOV64r0: return true;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4015 default: break;
4016 }
4017 OpcodeTablePtr = &RegOp2MemOpTable0;
4018 } else if (OpNum == 1) {
4019 OpcodeTablePtr = &RegOp2MemOpTable1;
4020 } else if (OpNum == 2) {
4021 OpcodeTablePtr = &RegOp2MemOpTable2;
Craig Topper7573c8f2012-08-31 22:12:16 +0000[diff] [blame]4022 } else if (OpNum == 3) {
4023 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4024 }
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4025
Chris Lattner626656a2010-10-08 03:54:52 +0000[diff] [blame]4026 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4027 return true;
Jakob Stoklund Olesen7a7b55e2010-07-09 20:43:13 +0000[diff] [blame]4028 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4029}
4030
4031bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4032 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4033 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4034 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4035 MemOp2RegOpTable.find(MI->getOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4036 if (I == MemOp2RegOpTable.end())
4037 return false;
4038 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4039 unsigned Index = I->second.second & TB_INDEX_MASK;
4040 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4041 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4042 if (UnfoldLoad && !FoldedLoad)
4043 return false;
4044 UnfoldLoad &= FoldedLoad;
4045 if (UnfoldStore && !FoldedStore)
4046 return false;
4047 UnfoldStore &= FoldedStore;
4048
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4049 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4050 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4051 if (!MI->hasOneMemOperand() &&
4052 RC == &X86::VR128RegClass &&
4053 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4054 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4055 // conservatively assume the address is unaligned. That's bad for
4056 // performance.
4057 return false;
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4058 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4059 SmallVector<MachineOperand,2> BeforeOps;
4060 SmallVector<MachineOperand,2> AfterOps;
4061 SmallVector<MachineOperand,4> ImpOps;
4062 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4063 MachineOperand &Op = MI->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4064 if (i >= Index && i < Index + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4065 AddrOps.push_back(Op);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4066 else if (Op.isReg() && Op.isImplicit())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4067 ImpOps.push_back(Op);
4068 else if (i < Index)
4069 BeforeOps.push_back(Op);
4070 else if (i > Index)
4071 AfterOps.push_back(Op);
4072 }
4073
4074 // Emit the load instruction.
4075 if (UnfoldLoad) {
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4076 std::pair<MachineInstr::mmo_iterator,
4077 MachineInstr::mmo_iterator> MMOs =
4078 MF.extractLoadMemRefs(MI->memoperands_begin(),
4079 MI->memoperands_end());
4080 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4081 if (UnfoldStore) {
4082 // Address operands cannot be marked isKill.
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4083 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4084 MachineOperand &MO = NewMIs[0]->getOperand(i);
Dan Gohman0d1e9a82008-10-03 15:45:36 +0000[diff] [blame]4085 if (MO.isReg())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4086 MO.setIsKill(false);
4087 }
4088 }
4089 }
4090
4091 // Emit the data processing instruction.
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4092 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4093 MachineInstrBuilder MIB(DataMI);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4094
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4095 if (FoldedStore)
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4096 MIB.addReg(Reg, RegState::Define);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4097 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4098 MIB.addOperand(BeforeOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4099 if (FoldedLoad)
4100 MIB.addReg(Reg);
4101 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
Dan Gohman2af1f852009-02-18 05:45:50 +0000[diff] [blame]4102 MIB.addOperand(AfterOps[i]);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4103 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4104 MachineOperand &MO = ImpOps[i];
Bill Wendlingf7b83c72009-05-13 21:33:08 +0000[diff] [blame]4105 MIB.addReg(MO.getReg(),
4106 getDefRegState(MO.isDef()) |
4107 RegState::Implicit |
4108 getKillRegState(MO.isKill()) |
Evan Cheng0dc101b2009-06-30 08:49:04 +0000[diff] [blame]4109 getDeadRegState(MO.isDead()) |
4110 getUndefRegState(MO.isUndef()));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4111 }
4112 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4113 switch (DataMI->getOpcode()) {
4114 default: break;
4115 case X86::CMP64ri32:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4116 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4117 case X86::CMP32ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4118 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4119 case X86::CMP16ri:
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4120 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4121 case X86::CMP8ri: {
4122 MachineOperand &MO0 = DataMI->getOperand(0);
4123 MachineOperand &MO1 = DataMI->getOperand(1);
4124 if (MO1.getImm() == 0) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4125 unsigned NewOpc;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4126 switch (DataMI->getOpcode()) {
Craig Topper4bc3e5a2012-08-21 08:16:16 +0000[diff] [blame]4127 default: llvm_unreachable("Unreachable!");
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4128 case X86::CMP64ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4129 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4130 case X86::CMP32ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4131 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
Dan Gohmanf8bf6632010-05-18 21:54:15 +0000[diff] [blame]4132 case X86::CMP16ri8:
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4133 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4134 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4135 }
Chris Lattner59687512008-01-11 18:10:50 +0000[diff] [blame]4136 DataMI->setDesc(get(NewOpc));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4137 MO1.ChangeToRegister(MO0.getReg(), false);
4138 }
4139 }
4140 }
4141 NewMIs.push_back(DataMI);
4142
4143 // Emit the store instruction.
4144 if (UnfoldStore) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4145 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4146 std::pair<MachineInstr::mmo_iterator,
4147 MachineInstr::mmo_iterator> MMOs =
4148 MF.extractStoreMemRefs(MI->memoperands_begin(),
4149 MI->memoperands_end());
4150 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4151 }
4152
4153 return true;
4154}
4155
4156bool
4157X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
Bill Wendling27b508d2009-02-11 21:51:19 +0000[diff] [blame]4158 SmallVectorImpl<SDNode*> &NewNodes) const {
Dan Gohman17059682008-07-17 19:10:17 +0000[diff] [blame]4159 if (!N->isMachineOpcode())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4160 return false;
4161
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4162 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4163 MemOp2RegOpTable.find(N->getMachineOpcode());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4164 if (I == MemOp2RegOpTable.end())
4165 return false;
4166 unsigned Opc = I->second.first;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4167 unsigned Index = I->second.second & TB_INDEX_MASK;
4168 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4169 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4170 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4171 MachineFunction &MF = DAG.getMachineFunction();
4172 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4173 unsigned NumDefs = MCID.NumDefs;
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4174 std::vector<SDValue> AddrOps;
4175 std::vector<SDValue> BeforeOps;
4176 std::vector<SDValue> AfterOps;
Dale Johannesen9f3f72f2009-02-06 01:31:28 +0000[diff] [blame]4177 DebugLoc dl = N->getDebugLoc();
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4178 unsigned NumOps = N->getNumOperands();
Dan Gohman48b185d2009-09-25 20:36:54 +0000[diff] [blame]4179 for (unsigned i = 0; i != NumOps-1; ++i) {
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4180 SDValue Op = N->getOperand(i);
Chris Lattnerec536272010-07-08 22:41:28 +0000[diff] [blame]4181 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4182 AddrOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4183 else if (i < Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4184 BeforeOps.push_back(Op);
Dan Gohmancc329b52009-03-04 19:23:38 +0000[diff] [blame]4185 else if (i > Index-NumDefs)
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4186 AfterOps.push_back(Op);
4187 }
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4188 SDValue Chain = N->getOperand(NumOps-1);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4189 AddrOps.push_back(Chain);
4190
4191 // Emit the load instruction.
4192 SDNode *Load = 0;
4193 if (FoldedLoad) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4194 EVT VT = *RC->vt_begin();
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4195 std::pair<MachineInstr::mmo_iterator,
4196 MachineInstr::mmo_iterator> MMOs =
4197 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4198 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4199 if (!(*MMOs.first) &&
4200 RC == &X86::VR128RegClass &&
4201 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4202 // Do not introduce a slow unaligned load.
4203 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4204 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4205 bool isAligned = (*MMOs.first) &&
4206 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4207 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
4208 VT, MVT::Other, &AddrOps[0], AddrOps.size());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4209 NewNodes.push_back(Load);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4210
4211 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4212 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4213 }
4214
4215 // Emit the data processing instruction.
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4216 std::vector<EVT> VTs;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4217 const TargetRegisterClass *DstRC = 0;
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4218 if (MCID.getNumDefs() > 0) {
Jakob Stoklund Olesen3c52f022012-05-07 22:10:26 +0000[diff] [blame]4219 DstRC = getRegClass(MCID, 0, &RI, MF);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4220 VTs.push_back(*DstRC->vt_begin());
4221 }
4222 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
Owen Anderson53aa7a92009-08-10 22:56:29 +0000[diff] [blame]4223 EVT VT = N->getValueType(i);
Evan Cheng6cc775f2011-06-28 19:10:37 +0000[diff] [blame]4224 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4225 VTs.push_back(VT);
4226 }
4227 if (Load)
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4228 BeforeOps.push_back(SDValue(Load, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4229 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4230 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0],
4231 BeforeOps.size());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4232 NewNodes.push_back(NewNode);
4233
4234 // Emit the store instruction.
4235 if (FoldedStore) {
4236 AddrOps.pop_back();
Dan Gohman2ce6f2a2008-07-27 21:46:04 +0000[diff] [blame]4237 AddrOps.push_back(SDValue(NewNode, 0));
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4238 AddrOps.push_back(Chain);
Evan Chengf25ef4f2009-11-16 21:56:03 +0000[diff] [blame]4239 std::pair<MachineInstr::mmo_iterator,
4240 MachineInstr::mmo_iterator> MMOs =
4241 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4242 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng0ce84482010-07-02 20:36:18 +0000[diff] [blame]4243 if (!(*MMOs.first) &&
4244 RC == &X86::VR128RegClass &&
4245 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4246 // Do not introduce a slow unaligned store.
4247 return false;
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4248 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4249 bool isAligned = (*MMOs.first) &&
4250 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman32f71d72009-09-25 18:54:59 +0000[diff] [blame]4251 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4252 isAligned, TM),
4253 dl, MVT::Other,
4254 &AddrOps[0], AddrOps.size());
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4255 NewNodes.push_back(Store);
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4256
4257 // Preserve memory reference information.
Dan Gohmandd76bb22009-10-09 18:10:05 +0000[diff] [blame]4258 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4259 }
4260
4261 return true;
4262}
4263
4264unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4265 bool UnfoldLoad, bool UnfoldStore,
4266 unsigned *LoadRegIndex) const {
Chris Lattner1c090c02010-10-07 23:08:41 +0000[diff] [blame]4267 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4268 MemOp2RegOpTable.find(Opc);
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4269 if (I == MemOp2RegOpTable.end())
4270 return 0;
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4271 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4272 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4273 if (UnfoldLoad && !FoldedLoad)
4274 return 0;
4275 if (UnfoldStore && !FoldedStore)
4276 return 0;
Dan Gohman49fa51d2009-10-30 22:18:41 +0000[diff] [blame]4277 if (LoadRegIndex)
Bruno Cardoso Lopes23eb5262011-09-08 18:35:57 +0000[diff] [blame]4278 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
Owen Anderson2a3be7b2008-01-07 01:35:02 +0000[diff] [blame]4279 return I->second.first;
4280}
4281
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4282bool
4283X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4284 int64_t &Offset1, int64_t &Offset2) const {
4285 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4286 return false;
4287 unsigned Opc1 = Load1->getMachineOpcode();
4288 unsigned Opc2 = Load2->getMachineOpcode();
4289 switch (Opc1) {
4290 default: return false;
4291 case X86::MOV8rm:
4292 case X86::MOV16rm:
4293 case X86::MOV32rm:
4294 case X86::MOV64rm:
4295 case X86::LD_Fp32m:
4296 case X86::LD_Fp64m:
4297 case X86::LD_Fp80m:
4298 case X86::MOVSSrm:
4299 case X86::MOVSDrm:
4300 case X86::MMX_MOVD64rm:
4301 case X86::MMX_MOVQ64rm:
4302 case X86::FsMOVAPSrm:
4303 case X86::FsMOVAPDrm:
4304 case X86::MOVAPSrm:
4305 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4306 case X86::MOVAPDrm:
4307 case X86::MOVDQArm:
4308 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4309 // AVX load instructions
4310 case X86::VMOVSSrm:
4311 case X86::VMOVSDrm:
4312 case X86::FsVMOVAPSrm:
4313 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4314 case X86::VMOVAPSrm:
4315 case X86::VMOVUPSrm:
4316 case X86::VMOVAPDrm:
4317 case X86::VMOVDQArm:
4318 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4319 case X86::VMOVAPSYrm:
4320 case X86::VMOVUPSYrm:
4321 case X86::VMOVAPDYrm:
4322 case X86::VMOVDQAYrm:
4323 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4324 break;
4325 }
4326 switch (Opc2) {
4327 default: return false;
4328 case X86::MOV8rm:
4329 case X86::MOV16rm:
4330 case X86::MOV32rm:
4331 case X86::MOV64rm:
4332 case X86::LD_Fp32m:
4333 case X86::LD_Fp64m:
4334 case X86::LD_Fp80m:
4335 case X86::MOVSSrm:
4336 case X86::MOVSDrm:
4337 case X86::MMX_MOVD64rm:
4338 case X86::MMX_MOVQ64rm:
4339 case X86::FsMOVAPSrm:
4340 case X86::FsMOVAPDrm:
4341 case X86::MOVAPSrm:
4342 case X86::MOVUPSrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4343 case X86::MOVAPDrm:
4344 case X86::MOVDQArm:
4345 case X86::MOVDQUrm:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4346 // AVX load instructions
4347 case X86::VMOVSSrm:
4348 case X86::VMOVSDrm:
4349 case X86::FsVMOVAPSrm:
4350 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopesd560b8c2011-09-14 02:36:58 +0000[diff] [blame]4351 case X86::VMOVAPSrm:
4352 case X86::VMOVUPSrm:
4353 case X86::VMOVAPDrm:
4354 case X86::VMOVDQArm:
4355 case X86::VMOVDQUrm:
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4356 case X86::VMOVAPSYrm:
4357 case X86::VMOVUPSYrm:
4358 case X86::VMOVAPDYrm:
4359 case X86::VMOVDQAYrm:
4360 case X86::VMOVDQUYrm:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4361 break;
4362 }
4363
4364 // Check if chain operands and base addresses match.
4365 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4366 Load1->getOperand(5) != Load2->getOperand(5))
4367 return false;
4368 // Segment operands should match as well.
4369 if (Load1->getOperand(4) != Load2->getOperand(4))
4370 return false;
4371 // Scale should be 1, Index should be Reg0.
4372 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4373 Load1->getOperand(2) == Load2->getOperand(2)) {
4374 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4375 return false;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4376
4377 // Now let's examine the displacements.
4378 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4379 isa<ConstantSDNode>(Load2->getOperand(3))) {
4380 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4381 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4382 return true;
4383 }
4384 }
4385 return false;
4386}
4387
4388bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4389 int64_t Offset1, int64_t Offset2,
4390 unsigned NumLoads) const {
4391 assert(Offset2 > Offset1);
4392 if ((Offset2 - Offset1) / 8 > 64)
4393 return false;
4394
4395 unsigned Opc1 = Load1->getMachineOpcode();
4396 unsigned Opc2 = Load2->getMachineOpcode();
4397 if (Opc1 != Opc2)
4398 return false; // FIXME: overly conservative?
4399
4400 switch (Opc1) {
4401 default: break;
4402 case X86::LD_Fp32m:
4403 case X86::LD_Fp64m:
4404 case X86::LD_Fp80m:
4405 case X86::MMX_MOVD64rm:
4406 case X86::MMX_MOVQ64rm:
4407 return false;
4408 }
4409
4410 EVT VT = Load1->getValueType(0);
4411 switch (VT.getSimpleVT().SimpleTy) {
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4412 default:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4413 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4414 // have 16 of them to play with.
4415 if (TM.getSubtargetImpl()->is64Bit()) {
4416 if (NumLoads >= 3)
4417 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4418 } else if (NumLoads) {
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4419 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4420 }
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4421 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4422 case MVT::i8:
4423 case MVT::i16:
4424 case MVT::i32:
4425 case MVT::i64:
Evan Cheng16cf9342010-01-22 23:49:11 +0000[diff] [blame]4426 case MVT::f32:
4427 case MVT::f64:
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4428 if (NumLoads)
4429 return false;
Bill Wendling8ce69cd2010-06-22 22:16:17 +0000[diff] [blame]4430 break;
Evan Cheng4f026f32010-01-22 03:34:51 +0000[diff] [blame]4431 }
4432
4433 return true;
4434}
4435
4436
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4437bool X86InstrInfo::
Owen Anderson4f6bf042008-08-14 22:49:33 +0000[diff] [blame]4438ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4439 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4440 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
Dan Gohman97d95d62008-10-21 03:29:32 +0000[diff] [blame]4441 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4442 return true;
Evan Chengf93bc7f2008-08-29 23:21:31 +0000[diff] [blame]4443 Cond[0].setImm(GetOppositeBranchCondition(CC));
Chris Lattner3a897f32006-10-21 05:52:40 +0000[diff] [blame]4444 return false;
Chris Lattnerc0fb5672006-10-20 17:42:20 +0000[diff] [blame]4445}
4446
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4447bool X86InstrInfo::
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4448isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4449 // FIXME: Return false for x87 stack register classes for now. We can't
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4450 // allow any loads of these registers before FpGet_ST0_80.
Evan Chengb5f0ec32009-02-06 17:17:30 +0000[diff] [blame]4451 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4452 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
Evan Chengf7137222008-10-27 07:14:50 +0000[diff] [blame]4453}
4454
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4455/// getGlobalBaseReg - Return a virtual register initialized with the
4456/// the global base register value. Output instructions required to
4457/// initialize the register in the function entry block, if necessary.
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4458///
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4459/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4460///
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4461unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4462 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4463 "X86-64 PIC uses RIP relative addressing");
4464
4465 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4466 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4467 if (GlobalBaseReg != 0)
4468 return GlobalBaseReg;
4469
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4470 // Create the register. The code to initialize it is inserted
4471 // later, by the CGBR pass (below).
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4472 MachineRegisterInfo &RegInfo = MF->getRegInfo();
Jakob Stoklund Olesen38dcd592012-05-20 18:43:00 +0000[diff] [blame]4473 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Dan Gohman6ebe7342008-09-30 00:58:23 +0000[diff] [blame]4474 X86FI->setGlobalBaseReg(GlobalBaseReg);
4475 return GlobalBaseReg;
Dan Gohman24300732008-09-23 18:22:58 +0000[diff] [blame]4476}
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4477
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4478// These are the replaceable SSE instructions. Some of these have Int variants
4479// that we don't include here. We don't want to replace instructions selected
4480// by intrinsics.
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4481static const uint16_t ReplaceableInstrs[][3] = {
Bruno Cardoso Lopes1401e042010-08-12 02:08:52 +0000[diff] [blame]4482 //PackedSingle PackedDouble PackedInt
Jakob Stoklund Olesendbff4e82010-03-30 22:46:53 +0000[diff] [blame]4483 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4484 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4485 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4486 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4487 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4488 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4489 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4490 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4491 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4492 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4493 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4494 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4495 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4496 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4497 // AVX 128-bit support
4498 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4499 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4500 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4501 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4502 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4503 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4504 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4505 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4506 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4507 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4508 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4509 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
Bruno Cardoso Lopes7f704b32010-08-12 20:20:53 +0000[diff] [blame]4510 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4511 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
Bruno Cardoso Lopes67785972011-07-14 18:50:58 +0000[diff] [blame]4512 // AVX 256-bit support
4513 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4514 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4515 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4516 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4517 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4518 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4519};
4520
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4521static const uint16_t ReplaceableInstrsAVX2[][3] = {
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4522 //PackedSingle PackedDouble PackedInt
Craig Topperf87a2be2011-11-09 09:37:21 +0000[diff] [blame]4523 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4524 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4525 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4526 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4527 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4528 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4529 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
Craig Topper12b72de2011-11-29 05:37:58 +0000[diff] [blame]4530 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4531 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4532 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4533 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4534 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4535 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4536 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4537};
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4538
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4539// FIXME: Some shuffle and unpack instructions have equivalents in different
4540// domains, but they require a bit more work than just switching opcodes.
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4541
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4542static const uint16_t *lookup(unsigned opcode, unsigned domain) {
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4543 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4544 if (ReplaceableInstrs[i][domain-1] == opcode)
4545 return ReplaceableInstrs[i];
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4546 return 0;
4547}
4548
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4549static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4550 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4551 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4552 return ReplaceableInstrsAVX2[i];
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4553 return 0;
4554}
4555
4556std::pair<uint16_t, uint16_t>
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]4557X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4558 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
Craig Topper05baa852011-11-15 05:55:35 +0000[diff] [blame]4559 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4560 uint16_t validDomains = 0;
4561 if (domain && lookup(MI->getOpcode(), domain))
4562 validDomains = 0xe;
4563 else if (domain && lookupAVX2(MI->getOpcode(), domain))
4564 validDomains = hasAVX2 ? 0xe : 0x6;
4565 return std::make_pair(domain, validDomains);
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4566}
4567
Jakob Stoklund Olesenb48c9942011-09-27 22:57:18 +0000[diff] [blame]4568void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4569 assert(Domain>0 && Domain<4 && "Invalid execution domain");
4570 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4571 assert(dom && "Not an SSE instruction");
Craig Topper2dac9622012-03-09 07:45:21 +0000[diff] [blame]4572 const uint16_t *table = lookup(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]4573 if (!table) { // try the other table
4574 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
4575 "256-bit vector operations only available in AVX2");
Craig Topper649d1c52011-11-15 06:39:01 +0000[diff] [blame]4576 table = lookupAVX2(MI->getOpcode(), dom);
Jakob Stoklund Olesen02845412011-11-23 04:03:08 +0000[diff] [blame]4577 }
Jakob Stoklund Olesenb551aa42010-03-29 23:24:21 +0000[diff] [blame]4578 assert(table && "Cannot change domain");
4579 MI->setDesc(get(table[Domain-1]));
Jakob Stoklund Olesen49e121d2010-03-25 17:25:00 +0000[diff] [blame]4580}
Chris Lattner6a5e7062010-04-26 23:37:21 +0000[diff] [blame]4581
4582/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
4583void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
4584 NopInst.setOpcode(X86::NOOP);
4585}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4586
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]4587bool X86InstrInfo::isHighLatencyDef(int opc) const {
4588 switch (opc) {
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4589 default: return false;
4590 case X86::DIVSDrm:
4591 case X86::DIVSDrm_Int:
4592 case X86::DIVSDrr:
4593 case X86::DIVSDrr_Int:
4594 case X86::DIVSSrm:
4595 case X86::DIVSSrm_Int:
4596 case X86::DIVSSrr:
4597 case X86::DIVSSrr_Int:
4598 case X86::SQRTPDm:
4599 case X86::SQRTPDm_Int:
4600 case X86::SQRTPDr:
4601 case X86::SQRTPDr_Int:
4602 case X86::SQRTPSm:
4603 case X86::SQRTPSm_Int:
4604 case X86::SQRTPSr:
4605 case X86::SQRTPSr_Int:
4606 case X86::SQRTSDm:
4607 case X86::SQRTSDm_Int:
4608 case X86::SQRTSDr:
4609 case X86::SQRTSDr_Int:
4610 case X86::SQRTSSm:
4611 case X86::SQRTSSm_Int:
4612 case X86::SQRTSSr:
4613 case X86::SQRTSSr_Int:
Bruno Cardoso Lopesc69d68a2011-09-15 22:15:52 +0000[diff] [blame]4614 // AVX instructions with high latency
4615 case X86::VDIVSDrm:
4616 case X86::VDIVSDrm_Int:
4617 case X86::VDIVSDrr:
4618 case X86::VDIVSDrr_Int:
4619 case X86::VDIVSSrm:
4620 case X86::VDIVSSrm_Int:
4621 case X86::VDIVSSrr:
4622 case X86::VDIVSSrr_Int:
4623 case X86::VSQRTPDm:
4624 case X86::VSQRTPDm_Int:
4625 case X86::VSQRTPDr:
4626 case X86::VSQRTPDr_Int:
4627 case X86::VSQRTPSm:
4628 case X86::VSQRTPSm_Int:
4629 case X86::VSQRTPSr:
4630 case X86::VSQRTPSr_Int:
4631 case X86::VSQRTSDm:
4632 case X86::VSQRTSDm_Int:
4633 case X86::VSQRTSDr:
4634 case X86::VSQRTSSm:
4635 case X86::VSQRTSSm_Int:
4636 case X86::VSQRTSSr:
Evan Cheng63c76082010-10-19 18:58:51 +0000[diff] [blame]4637 return true;
4638 }
4639}
4640
Andrew Trick641e2d42011-03-05 08:00:22 +0000[diff] [blame]4641bool X86InstrInfo::
4642hasHighOperandLatency(const InstrItineraryData *ItinData,
4643 const MachineRegisterInfo *MRI,
4644 const MachineInstr *DefMI, unsigned DefIdx,
4645 const MachineInstr *UseMI, unsigned UseIdx) const {
4646 return isHighLatencyDef(DefMI->getOpcode());
4647}
4648
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4649namespace {
4650 /// CGBR - Create Global Base Reg pass. This initializes the PIC
4651 /// global base register for x86-32.
4652 struct CGBR : public MachineFunctionPass {
4653 static char ID;
Owen Andersona7aed182010-08-06 18:33:48 +0000[diff] [blame]4654 CGBR() : MachineFunctionPass(ID) {}
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4655
4656 virtual bool runOnMachineFunction(MachineFunction &MF) {
4657 const X86TargetMachine *TM =
4658 static_cast<const X86TargetMachine *>(&MF.getTarget());
4659
4660 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
4661 "X86-64 PIC uses RIP relative addressing");
4662
4663 // Only emit a global base reg in PIC mode.
4664 if (TM->getRelocationModel() != Reloc::PIC_)
4665 return false;
4666
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]4667 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
4668 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4669
4670 // If we didn't need a GlobalBaseReg, don't insert code.
4671 if (GlobalBaseReg == 0)
4672 return false;
4673
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4674 // Insert the set of GlobalBaseReg into the first MBB of the function
4675 MachineBasicBlock &FirstMBB = MF.front();
4676 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
4677 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
4678 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4679 const X86InstrInfo *TII = TM->getInstrInfo();
4680
4681 unsigned PC;
4682 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
Craig Topperabadc662012-04-20 06:31:50 +0000[diff] [blame]4683 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4684 else
Dan Gohman534db8a2010-09-17 20:24:24 +0000[diff] [blame]4685 PC = GlobalBaseReg;
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4686
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4687 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4688 // only used in JIT code emission as displacement to pc.
4689 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
NAKAMURA Takumi9d29eff2011-01-26 02:03:37 +0000[diff] [blame]4690
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4691 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4692 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4693 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
Dan Gohmand7b5ce32010-07-10 09:00:22 +0000[diff] [blame]4694 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4695 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4696 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4697 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4698 }
4699
4700 return true;
4701 }
4702
4703 virtual const char *getPassName() const {
4704 return "X86 PIC Global Base Reg Initialization";
4705 }
4706
4707 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4708 AU.setPreservesCFG();
4709 MachineFunctionPass::getAnalysisUsage(AU);
4710 }
4711 };
4712}
4713
4714char CGBR::ID = 0;
4715FunctionPass*
4716llvm::createGlobalBaseRegPass() { return new CGBR(); }
Hans Wennborg789acfb2012-06-01 16:27:21 +0000[diff] [blame]4717
4718namespace {
4719 struct LDTLSCleanup : public MachineFunctionPass {
4720 static char ID;
4721 LDTLSCleanup() : MachineFunctionPass(ID) {}
4722
4723 virtual bool runOnMachineFunction(MachineFunction &MF) {
4724 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4725 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4726 // No point folding accesses if there isn't at least two.
4727 return false;
4728 }
4729
4730 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4731 return VisitNode(DT->getRootNode(), 0);
4732 }
4733
4734 // Visit the dominator subtree rooted at Node in pre-order.
4735 // If TLSBaseAddrReg is non-null, then use that to replace any
4736 // TLS_base_addr instructions. Otherwise, create the register
4737 // when the first such instruction is seen, and then use it
4738 // as we encounter more instructions.
4739 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4740 MachineBasicBlock *BB = Node->getBlock();
4741 bool Changed = false;
4742
4743 // Traverse the current block.
4744 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4745 ++I) {
4746 switch (I->getOpcode()) {
4747 case X86::TLS_base_addr32:
4748 case X86::TLS_base_addr64:
4749 if (TLSBaseAddrReg)
4750 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4751 else
4752 I = SetRegister(I, &TLSBaseAddrReg);
4753 Changed = true;
4754 break;
4755 default:
4756 break;
4757 }
4758 }
4759
4760 // Visit the children of this block in the dominator tree.
4761 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4762 I != E; ++I) {
4763 Changed |= VisitNode(*I, TLSBaseAddrReg);
4764 }
4765
4766 return Changed;
4767 }
4768
4769 // Replace the TLS_base_addr instruction I with a copy from
4770 // TLSBaseAddrReg, returning the new instruction.
4771 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4772 unsigned TLSBaseAddrReg) {
4773 MachineFunction *MF = I->getParent()->getParent();
4774 const X86TargetMachine *TM =
4775 static_cast<const X86TargetMachine *>(&MF->getTarget());
4776 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4777 const X86InstrInfo *TII = TM->getInstrInfo();
4778
4779 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4780 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4781 TII->get(TargetOpcode::COPY),
4782 is64Bit ? X86::RAX : X86::EAX)
4783 .addReg(TLSBaseAddrReg);
4784
4785 // Erase the TLS_base_addr instruction.
4786 I->eraseFromParent();
4787
4788 return Copy;
4789 }
4790
4791 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4792 // inserting a copy instruction after I. Returns the new instruction.
4793 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4794 MachineFunction *MF = I->getParent()->getParent();
4795 const X86TargetMachine *TM =
4796 static_cast<const X86TargetMachine *>(&MF->getTarget());
4797 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4798 const X86InstrInfo *TII = TM->getInstrInfo();
4799
4800 // Create a virtual register for the TLS base address.
4801 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4802 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4803 ? &X86::GR64RegClass
4804 : &X86::GR32RegClass);
4805
4806 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4807 MachineInstr *Next = I->getNextNode();
4808 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4809 TII->get(TargetOpcode::COPY),
4810 *TLSBaseAddrReg)
4811 .addReg(is64Bit ? X86::RAX : X86::EAX);
4812
4813 return Copy;
4814 }
4815
4816 virtual const char *getPassName() const {
4817 return "Local Dynamic TLS Access Clean-up";
4818 }
4819
4820 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4821 AU.setPreservesCFG();
4822 AU.addRequired<MachineDominatorTree>();
4823 MachineFunctionPass::getAnalysisUsage(AU);
4824 }
4825 };
4826}
4827
4828char LDTLSCleanup::ID = 0;
4829FunctionPass*
4830llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }