Gitiles
Code Review Sign In
gerrit-public.fairphone.software / fp2-dev / platform / external / llvm / e6ac7d69d3308d580a87ff3e3645f98e2bc565a5 / . / lib / Target / X86 / X86InstrInfo.cpp
blob: 4f3d824b4a7638c6cbeeb5fd279e09eae6ff91a0 [file] [log] [blame]
Jia Liu31d157a2012-02-18 12:03:15 +0000[diff] [blame]1//===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
Misha Brukman0e0a7a452005-04-21 23:38:14 +0000[diff] [blame]2//
John Criswellb576c942003-10-20 19:43:21 +0000[diff] [blame]3// The LLVM Compiler Infrastructure
4//
Chris Lattner4ee451d2007-12-29 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 Brukman0e0a7a452005-04-21 23:38:14 +0000[diff] [blame]7//
John Criswellb576c942003-10-20 19:43:21 +0000[diff] [blame]8//===----------------------------------------------------------------------===//
Chris Lattner72614082002-10-25 22:55:53 +0000[diff] [blame]9//
Chris Lattner3501fea2003-01-14 22:00:31 +0000[diff] [blame]10// This file contains the X86 implementation of the TargetInstrInfo class.
Chris Lattner72614082002-10-25 22:55:53 +0000[diff] [blame]11//
12//===----------------------------------------------------------------------===//
13
Chris Lattner055c9652002-10-29 21:05:24 +0000[diff] [blame]14#include "X86InstrInfo.h"
Chris Lattner4ce42a72002-12-03 05:42:53 +0000[diff] [blame]15#include "X86.h"
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]16#include "X86InstrBuilder.h"
Owen Andersond94b6a12008-01-04 23:57:37 +0000[diff] [blame]17#include "X86MachineFunctionInfo.h"
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]18#include "X86Subtarget.h"
19#include "X86TargetMachine.h"
Dan Gohmand68a0762009-01-05 17:59:02 +0000[diff] [blame]20#include "llvm/DerivedTypes.h"
Owen Anderson0a5372e2009-07-13 04:09:18 +0000[diff] [blame]21#include "llvm/LLVMContext.h"
Owen Anderson718cb662007-09-07 04:06:50 +0000[diff] [blame]22#include "llvm/ADT/STLExtras.h"
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]23#include "llvm/CodeGen/MachineConstantPool.h"
Hans Wennborgf0234fc2012-06-01 16:27:21 +0000[diff] [blame]24#include "llvm/CodeGen/MachineDominators.h"
Owen Andersond94b6a12008-01-04 23:57:37 +0000[diff] [blame]25#include "llvm/CodeGen/MachineFrameInfo.h"
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]26#include "llvm/CodeGen/MachineInstrBuilder.h"
Chris Lattner84bc5422007-12-31 04:13:23 +0000[diff] [blame]27#include "llvm/CodeGen/MachineRegisterInfo.h"
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]28#include "llvm/CodeGen/LiveVariables.h"
Craig Topper79aa3412012-03-17 18:46:09 +0000[diff] [blame]29#include "llvm/MC/MCAsmInfo.h"
Chris Lattneree9eb412010-04-26 23:37:21 +0000[diff] [blame]30#include "llvm/MC/MCInst.h"
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]31#include "llvm/Support/CommandLine.h"
David Greene5b901322010-01-05 01:29:29 +0000[diff] [blame]32#include "llvm/Support/Debug.h"
Torok Edwinab7c09b2009-07-08 18:01:40 +0000[diff] [blame]33#include "llvm/Support/ErrorHandling.h"
34#include "llvm/Support/raw_ostream.h"
Evan Cheng0488db92007-09-25 01:57:46 +0000[diff] [blame]35#include "llvm/Target/TargetOptions.h"
David Greeneb87bc952009-11-12 20:55:29 +0000[diff] [blame]36#include <limits>
37
Evan Cheng4db3cff2011-07-01 17:57:27 +0000[diff] [blame]38#define GET_INSTRINFO_CTOR
Evan Cheng22fee2d2011-06-28 20:07:07 +0000[diff] [blame]39#include "X86GenInstrInfo.inc"
40
Brian Gaeked0fde302003-11-11 22:41:34 +0000[diff] [blame]41using namespace llvm;
42
Chris Lattner705e07f2009-08-23 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 Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]55
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]56enum {
57 // Select which memory operand is being unfolded.
Craig Topper3ed920f2012-06-23 08:01:18 +0000[diff] [blame]58 // (stored in bits 0 - 3)
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]59 TB_INDEX_0 = 0,
60 TB_INDEX_1 = 1,
61 TB_INDEX_2 = 2,
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]62 TB_INDEX_3 = 3,
Craig Topper3ed920f2012-06-23 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 Lopescbf479d2011-09-08 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 Topper3ed920f2012-06-23 08:01:18 +0000[diff] [blame]84 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]85};
86
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]87struct X86OpTblEntry {
88 uint16_t RegOp;
89 uint16_t MemOp;
Craig Topper3ed920f2012-06-23 08:01:18 +0000[diff] [blame]90 uint16_t Flags;
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]91};
92
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]93X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
Evan Cheng4db3cff2011-07-01 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 Cheng25ab6902006-09-08 06:48:29 +0000[diff] [blame]100 TM(tm), RI(tm, *this) {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]101
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]102 static const X86OpTblEntry OpTbl2Addr[] = {
Bruno Cardoso Lopescbf479d2011-09-08 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 Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]267 };
268
269 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
Craig Topper72051bf2012-03-09 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 Lopescbf479d2011-09-08 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 Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]277 }
278
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]279 static const X86OpTblEntry OpTbl0[] = {
Bruno Cardoso Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 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 Lopes484ddf52011-09-14 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 Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]361 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 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 Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]372 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 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 Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]378 };
379
380 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
Craig Topper72051bf2012-03-09 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 Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]384 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
385 RegOp, MemOp, TB_INDEX_0 | Flags);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]386 }
387
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]388 static const X86OpTblEntry OpTbl1[] = {
Bruno Cardoso Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]413 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
414 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
Craig Topper8e58b3e2012-06-15 07:02:58 +0000[diff] [blame]415 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
416 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
Bruno Cardoso Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 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 Lopescbf479d2011-09-08 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 Topperdcce2442011-11-14 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 Lopescbf479d2011-09-08 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 Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]481 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]482 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
483 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 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 Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]487 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
488 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
Craig Topper8e58b3e2012-06-15 07:02:58 +0000[diff] [blame]489 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
490 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
Pete Cooper312091e2012-06-14 22:12:58 +0000[diff] [blame]491 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
Craig Topper8e58b3e2012-06-15 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 Lopes484ddf52011-09-14 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 Topperdcce2442011-11-14 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 Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]521 { X86::VPERMILPDri, X86::VPERMILPDmi, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]522 { X86::VPERMILPSri, X86::VPERMILPSmi, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 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 Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]534 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]535 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
Nadav Rotemaec9f382012-07-15 12:26:30 +0000[diff] [blame]536 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
537
Bruno Cardoso Lopes484ddf52011-09-14 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 Topper40385c82012-01-19 08:50:38 +0000[diff] [blame]541 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]542 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]543 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
Craig Topper40385c82012-01-19 08:50:38 +0000[diff] [blame]544 { X86::VPERMILPDYri, X86::VPERMILPDYmi, TB_ALIGN_32 },
545 { X86::VPERMILPSYri, X86::VPERMILPSYmi, TB_ALIGN_32 },
Nadav Rotemaec9f382012-07-15 12:26:30 +0000[diff] [blame]546
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]547 // AVX2 foldable instructions
Craig Topper40385c82012-01-19 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 Rotemaec9f382012-07-15 12:26:30 +0000[diff] [blame]562 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
563 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]564 };
565
566 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]567 unsigned RegOp = OpTbl1[i].RegOp;
568 unsigned MemOp = OpTbl1[i].MemOp;
569 unsigned Flags = OpTbl1[i].Flags;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]570 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
571 RegOp, MemOp,
572 // Index 1, folded load
573 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]574 }
575
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]576 static const X86OpTblEntry OpTbl2[] = {
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]577 { X86::ADC32rr, X86::ADC32rm, 0 },
578 { X86::ADC64rr, X86::ADC64rm, 0 },
579 { X86::ADD16rr, X86::ADD16rm, 0 },
580 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
581 { X86::ADD32rr, X86::ADD32rm, 0 },
582 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
583 { X86::ADD64rr, X86::ADD64rm, 0 },
584 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
585 { X86::ADD8rr, X86::ADD8rm, 0 },
586 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
587 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
588 { X86::ADDSDrr, X86::ADDSDrm, 0 },
589 { X86::ADDSSrr, X86::ADDSSrm, 0 },
590 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
591 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
592 { X86::AND16rr, X86::AND16rm, 0 },
593 { X86::AND32rr, X86::AND32rm, 0 },
594 { X86::AND64rr, X86::AND64rm, 0 },
595 { X86::AND8rr, X86::AND8rm, 0 },
596 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
597 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
598 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
599 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]600 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
601 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
602 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
603 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]604 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
605 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
606 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
607 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
608 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
609 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
610 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
611 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
612 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
613 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
614 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
615 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
616 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
617 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
618 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
619 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
620 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
621 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
622 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
623 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
624 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
625 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
626 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
627 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
628 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
629 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
630 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
631 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
632 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
633 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
634 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
635 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
636 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
637 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
638 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
639 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
640 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
641 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
642 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
643 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
644 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
645 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
646 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
647 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
648 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
649 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
650 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
651 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
652 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
653 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
654 { X86::CMPSDrr, X86::CMPSDrm, 0 },
655 { X86::CMPSSrr, X86::CMPSSrm, 0 },
656 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
657 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
658 { X86::DIVSDrr, X86::DIVSDrm, 0 },
659 { X86::DIVSSrr, X86::DIVSSrm, 0 },
660 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
661 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
662 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
663 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
664 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
665 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
666 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
667 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
668 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
669 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
670 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
671 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
672 { X86::IMUL16rr, X86::IMUL16rm, 0 },
673 { X86::IMUL32rr, X86::IMUL32rm, 0 },
674 { X86::IMUL64rr, X86::IMUL64rm, 0 },
675 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
676 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
Manman Renc586d262012-08-13 18:29:41 +0000[diff] [blame]677 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
678 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
679 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
680 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
681 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
682 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]683 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
684 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, TB_ALIGN_16 },
685 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
686 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, TB_ALIGN_16 },
687 { X86::MAXSDrr, X86::MAXSDrm, 0 },
688 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
689 { X86::MAXSSrr, X86::MAXSSrm, 0 },
690 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
691 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
692 { X86::MINPDrr_Int, X86::MINPDrm_Int, TB_ALIGN_16 },
693 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
694 { X86::MINPSrr_Int, X86::MINPSrm_Int, TB_ALIGN_16 },
695 { X86::MINSDrr, X86::MINSDrm, 0 },
696 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
697 { X86::MINSSrr, X86::MINSSrm, 0 },
698 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]699 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]700 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
701 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
702 { X86::MULSDrr, X86::MULSDrm, 0 },
703 { X86::MULSSrr, X86::MULSSrm, 0 },
704 { X86::OR16rr, X86::OR16rm, 0 },
705 { X86::OR32rr, X86::OR32rm, 0 },
706 { X86::OR64rr, X86::OR64rm, 0 },
707 { X86::OR8rr, X86::OR8rm, 0 },
708 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
709 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
710 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
711 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]712 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]713 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
714 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
715 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
716 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
717 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
718 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]719 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
720 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]721 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]722 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]723 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
724 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
725 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
726 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]727 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]728 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
729 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]730 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]731 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
732 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
733 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]734 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]735 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]736 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
737 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]738 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]739 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]740 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]741 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]742 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]743 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]744 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
745 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
746 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
747 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
748 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
749 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]750 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]751 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
752 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
753 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
754 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
755 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
756 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
757 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
Craig Topper969ba282012-01-25 06:43:11 +0000[diff] [blame]758 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
759 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
760 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
761 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]762 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
763 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
764 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
765 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
766 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
767 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
768 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
769 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
770 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
771 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
772 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
773 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
774 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
775 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
776 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
777 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
778 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
779 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
780 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
781 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
782 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
783 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
784 { X86::SBB32rr, X86::SBB32rm, 0 },
785 { X86::SBB64rr, X86::SBB64rm, 0 },
786 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
787 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
788 { X86::SUB16rr, X86::SUB16rm, 0 },
789 { X86::SUB32rr, X86::SUB32rm, 0 },
790 { X86::SUB64rr, X86::SUB64rm, 0 },
791 { X86::SUB8rr, X86::SUB8rm, 0 },
792 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
793 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
794 { X86::SUBSDrr, X86::SUBSDrm, 0 },
795 { X86::SUBSSrr, X86::SUBSSrm, 0 },
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]796 // FIXME: TEST*rr -> swapped operand of TEST*mr.
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]797 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
798 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
799 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
800 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
801 { X86::XOR16rr, X86::XOR16rm, 0 },
802 { X86::XOR32rr, X86::XOR32rm, 0 },
803 { X86::XOR64rr, X86::XOR64rm, 0 },
804 { X86::XOR8rr, X86::XOR8rm, 0 },
805 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]806 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
807 // AVX 128-bit versions of foldable instructions
808 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
809 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
810 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
811 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
812 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
813 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
814 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
815 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
816 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
817 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
818 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
819 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
Craig Topper13d89c72012-06-25 06:16:00 +0000[diff] [blame]820 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]821 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, TB_ALIGN_16 },
822 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
823 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
824 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
825 { X86::VADDPDrr, X86::VADDPDrm, TB_ALIGN_16 },
826 { X86::VADDPSrr, X86::VADDPSrm, TB_ALIGN_16 },
827 { X86::VADDSDrr, X86::VADDSDrm, 0 },
828 { X86::VADDSSrr, X86::VADDSSrm, 0 },
829 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, TB_ALIGN_16 },
830 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, TB_ALIGN_16 },
831 { X86::VANDNPDrr, X86::VANDNPDrm, TB_ALIGN_16 },
832 { X86::VANDNPSrr, X86::VANDNPSrm, TB_ALIGN_16 },
833 { X86::VANDPDrr, X86::VANDPDrm, TB_ALIGN_16 },
834 { X86::VANDPSrr, X86::VANDPSrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]835 { X86::VBLENDPDrri, X86::VBLENDPDrmi, TB_ALIGN_16 },
836 { X86::VBLENDPSrri, X86::VBLENDPSrmi, TB_ALIGN_16 },
837 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, TB_ALIGN_16 },
838 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]839 { X86::VCMPPDrri, X86::VCMPPDrmi, TB_ALIGN_16 },
840 { X86::VCMPPSrri, X86::VCMPPSrmi, TB_ALIGN_16 },
841 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
842 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
843 { X86::VDIVPDrr, X86::VDIVPDrm, TB_ALIGN_16 },
844 { X86::VDIVPSrr, X86::VDIVPSrm, TB_ALIGN_16 },
845 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
846 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
847 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
848 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
849 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
850 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
851 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
852 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
853 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
854 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
855 { X86::VHADDPDrr, X86::VHADDPDrm, TB_ALIGN_16 },
856 { X86::VHADDPSrr, X86::VHADDPSrm, TB_ALIGN_16 },
857 { X86::VHSUBPDrr, X86::VHSUBPDrm, TB_ALIGN_16 },
858 { X86::VHSUBPSrr, X86::VHSUBPSrm, TB_ALIGN_16 },
859 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
860 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
861 { X86::VMAXPDrr, X86::VMAXPDrm, TB_ALIGN_16 },
862 { X86::VMAXPDrr_Int, X86::VMAXPDrm_Int, TB_ALIGN_16 },
863 { X86::VMAXPSrr, X86::VMAXPSrm, TB_ALIGN_16 },
864 { X86::VMAXPSrr_Int, X86::VMAXPSrm_Int, TB_ALIGN_16 },
865 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
866 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 },
867 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
868 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 },
869 { X86::VMINPDrr, X86::VMINPDrm, TB_ALIGN_16 },
870 { X86::VMINPDrr_Int, X86::VMINPDrm_Int, TB_ALIGN_16 },
871 { X86::VMINPSrr, X86::VMINPSrm, TB_ALIGN_16 },
872 { X86::VMINPSrr_Int, X86::VMINPSrm_Int, TB_ALIGN_16 },
873 { X86::VMINSDrr, X86::VMINSDrm, 0 },
874 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 },
875 { X86::VMINSSrr, X86::VMINSSrm, 0 },
876 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]877 { X86::VMPSADBWrri, X86::VMPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]878 { X86::VMULPDrr, X86::VMULPDrm, TB_ALIGN_16 },
879 { X86::VMULPSrr, X86::VMULPSrm, TB_ALIGN_16 },
880 { X86::VMULSDrr, X86::VMULSDrm, 0 },
881 { X86::VMULSSrr, X86::VMULSSrm, 0 },
882 { X86::VORPDrr, X86::VORPDrm, TB_ALIGN_16 },
883 { X86::VORPSrr, X86::VORPSrm, TB_ALIGN_16 },
884 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, TB_ALIGN_16 },
885 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]886 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]887 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, TB_ALIGN_16 },
888 { X86::VPADDBrr, X86::VPADDBrm, TB_ALIGN_16 },
889 { X86::VPADDDrr, X86::VPADDDrm, TB_ALIGN_16 },
890 { X86::VPADDQrr, X86::VPADDQrm, TB_ALIGN_16 },
891 { X86::VPADDSBrr, X86::VPADDSBrm, TB_ALIGN_16 },
892 { X86::VPADDSWrr, X86::VPADDSWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]893 { X86::VPADDUSBrr, X86::VPADDUSBrm, TB_ALIGN_16 },
894 { X86::VPADDUSWrr, X86::VPADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]895 { X86::VPADDWrr, X86::VPADDWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]896 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]897 { X86::VPANDNrr, X86::VPANDNrm, TB_ALIGN_16 },
898 { X86::VPANDrr, X86::VPANDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]899 { X86::VPAVGBrr, X86::VPAVGBrm, TB_ALIGN_16 },
900 { X86::VPAVGWrr, X86::VPAVGWrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]901 { X86::VPBLENDWrri, X86::VPBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]902 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, TB_ALIGN_16 },
903 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]904 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]905 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, TB_ALIGN_16 },
906 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, TB_ALIGN_16 },
907 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]908 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]909 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]910 { X86::VPHADDDrr, X86::VPHADDDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]911 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]912 { X86::VPHADDWrr, X86::VPHADDWrm, TB_ALIGN_16 },
913 { X86::VPHSUBDrr, X86::VPHSUBDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]914 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]915 { X86::VPHSUBWrr, X86::VPHSUBWrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]916 { X86::VPERMILPDrr, X86::VPERMILPDrm, TB_ALIGN_16 },
917 { X86::VPERMILPSrr, X86::VPERMILPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]918 { X86::VPINSRWrri, X86::VPINSRWrmi, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]919 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]920 { X86::VPMADDWDrr, X86::VPMADDWDrm, TB_ALIGN_16 },
921 { X86::VPMAXSWrr, X86::VPMAXSWrm, TB_ALIGN_16 },
922 { X86::VPMAXUBrr, X86::VPMAXUBrm, TB_ALIGN_16 },
923 { X86::VPMINSWrr, X86::VPMINSWrm, TB_ALIGN_16 },
924 { X86::VPMINUBrr, X86::VPMINUBrm, TB_ALIGN_16 },
925 { X86::VPMULDQrr, X86::VPMULDQrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]926 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]927 { X86::VPMULHUWrr, X86::VPMULHUWrm, TB_ALIGN_16 },
928 { X86::VPMULHWrr, X86::VPMULHWrm, TB_ALIGN_16 },
929 { X86::VPMULLDrr, X86::VPMULLDrm, TB_ALIGN_16 },
930 { X86::VPMULLWrr, X86::VPMULLWrm, TB_ALIGN_16 },
931 { X86::VPMULUDQrr, X86::VPMULUDQrm, TB_ALIGN_16 },
932 { X86::VPORrr, X86::VPORrm, TB_ALIGN_16 },
933 { X86::VPSADBWrr, X86::VPSADBWrm, TB_ALIGN_16 },
Craig Topper969ba282012-01-25 06:43:11 +0000[diff] [blame]934 { X86::VPSHUFBrr, X86::VPSHUFBrm, TB_ALIGN_16 },
935 { X86::VPSIGNBrr, X86::VPSIGNBrm, TB_ALIGN_16 },
936 { X86::VPSIGNWrr, X86::VPSIGNWrm, TB_ALIGN_16 },
937 { X86::VPSIGNDrr, X86::VPSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]938 { X86::VPSLLDrr, X86::VPSLLDrm, TB_ALIGN_16 },
939 { X86::VPSLLQrr, X86::VPSLLQrm, TB_ALIGN_16 },
940 { X86::VPSLLWrr, X86::VPSLLWrm, TB_ALIGN_16 },
941 { X86::VPSRADrr, X86::VPSRADrm, TB_ALIGN_16 },
942 { X86::VPSRAWrr, X86::VPSRAWrm, TB_ALIGN_16 },
943 { X86::VPSRLDrr, X86::VPSRLDrm, TB_ALIGN_16 },
944 { X86::VPSRLQrr, X86::VPSRLQrm, TB_ALIGN_16 },
945 { X86::VPSRLWrr, X86::VPSRLWrm, TB_ALIGN_16 },
946 { X86::VPSUBBrr, X86::VPSUBBrm, TB_ALIGN_16 },
947 { X86::VPSUBDrr, X86::VPSUBDrm, TB_ALIGN_16 },
948 { X86::VPSUBSBrr, X86::VPSUBSBrm, TB_ALIGN_16 },
949 { X86::VPSUBSWrr, X86::VPSUBSWrm, TB_ALIGN_16 },
950 { X86::VPSUBWrr, X86::VPSUBWrm, TB_ALIGN_16 },
951 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, TB_ALIGN_16 },
952 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, TB_ALIGN_16 },
953 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, TB_ALIGN_16 },
954 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, TB_ALIGN_16 },
955 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, TB_ALIGN_16 },
956 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, TB_ALIGN_16 },
957 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, TB_ALIGN_16 },
958 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, TB_ALIGN_16 },
959 { X86::VPXORrr, X86::VPXORrm, TB_ALIGN_16 },
960 { X86::VSHUFPDrri, X86::VSHUFPDrmi, TB_ALIGN_16 },
961 { X86::VSHUFPSrri, X86::VSHUFPSrmi, TB_ALIGN_16 },
962 { X86::VSUBPDrr, X86::VSUBPDrm, TB_ALIGN_16 },
963 { X86::VSUBPSrr, X86::VSUBPSrm, TB_ALIGN_16 },
964 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
965 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
966 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, TB_ALIGN_16 },
967 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, TB_ALIGN_16 },
968 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, TB_ALIGN_16 },
969 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, TB_ALIGN_16 },
970 { X86::VXORPDrr, X86::VXORPDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]971 { X86::VXORPSrr, X86::VXORPSrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]972 // AVX 256-bit foldable instructions
973 { X86::VADDPDYrr, X86::VADDPDYrm, TB_ALIGN_32 },
974 { X86::VADDPSYrr, X86::VADDPSYrm, TB_ALIGN_32 },
975 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, TB_ALIGN_32 },
976 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, TB_ALIGN_32 },
977 { X86::VANDNPDYrr, X86::VANDNPDYrm, TB_ALIGN_32 },
978 { X86::VANDNPSYrr, X86::VANDNPSYrm, TB_ALIGN_32 },
979 { X86::VANDPDYrr, X86::VANDPDYrm, TB_ALIGN_32 },
980 { X86::VANDPSYrr, X86::VANDPSYrm, TB_ALIGN_32 },
981 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, TB_ALIGN_32 },
982 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, TB_ALIGN_32 },
983 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, TB_ALIGN_32 },
984 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, TB_ALIGN_32 },
985 { X86::VCMPPDYrri, X86::VCMPPDYrmi, TB_ALIGN_32 },
986 { X86::VCMPPSYrri, X86::VCMPPSYrmi, TB_ALIGN_32 },
987 { X86::VDIVPDYrr, X86::VDIVPDYrm, TB_ALIGN_32 },
988 { X86::VDIVPSYrr, X86::VDIVPSYrm, TB_ALIGN_32 },
989 { X86::VHADDPDYrr, X86::VHADDPDYrm, TB_ALIGN_32 },
990 { X86::VHADDPSYrr, X86::VHADDPSYrm, TB_ALIGN_32 },
991 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, TB_ALIGN_32 },
992 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, TB_ALIGN_32 },
993 { X86::VINSERTF128rr, X86::VINSERTF128rm, TB_ALIGN_32 },
994 { X86::VMAXPDYrr, X86::VMAXPDYrm, TB_ALIGN_32 },
995 { X86::VMAXPDYrr_Int, X86::VMAXPDYrm_Int, TB_ALIGN_32 },
996 { X86::VMAXPSYrr, X86::VMAXPSYrm, TB_ALIGN_32 },
997 { X86::VMAXPSYrr_Int, X86::VMAXPSYrm_Int, TB_ALIGN_32 },
998 { X86::VMINPDYrr, X86::VMINPDYrm, TB_ALIGN_32 },
999 { X86::VMINPDYrr_Int, X86::VMINPDYrm_Int, TB_ALIGN_32 },
1000 { X86::VMINPSYrr, X86::VMINPSYrm, TB_ALIGN_32 },
1001 { X86::VMINPSYrr_Int, X86::VMINPSYrm_Int, TB_ALIGN_32 },
1002 { X86::VMULPDYrr, X86::VMULPDYrm, TB_ALIGN_32 },
1003 { X86::VMULPSYrr, X86::VMULPSYrm, TB_ALIGN_32 },
1004 { X86::VORPDYrr, X86::VORPDYrm, TB_ALIGN_32 },
1005 { X86::VORPSYrr, X86::VORPSYrm, TB_ALIGN_32 },
1006 { X86::VPERM2F128rr, X86::VPERM2F128rm, TB_ALIGN_32 },
1007 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, TB_ALIGN_32 },
1008 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, TB_ALIGN_32 },
1009 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, TB_ALIGN_32 },
1010 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, TB_ALIGN_32 },
1011 { X86::VSUBPDYrr, X86::VSUBPDYrm, TB_ALIGN_32 },
1012 { X86::VSUBPSYrr, X86::VSUBPSYrm, TB_ALIGN_32 },
1013 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, TB_ALIGN_32 },
1014 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, TB_ALIGN_32 },
1015 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, TB_ALIGN_32 },
1016 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, TB_ALIGN_32 },
1017 { X86::VXORPDYrr, X86::VXORPDYrm, TB_ALIGN_32 },
1018 { X86::VXORPSYrr, X86::VXORPSYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1019 // AVX2 foldable instructions
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1020 { X86::VINSERTI128rr, X86::VINSERTI128rm, TB_ALIGN_16 },
1021 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, TB_ALIGN_32 },
1022 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, TB_ALIGN_32 },
1023 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, TB_ALIGN_32 },
1024 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, TB_ALIGN_32 },
1025 { X86::VPADDBYrr, X86::VPADDBYrm, TB_ALIGN_32 },
1026 { X86::VPADDDYrr, X86::VPADDDYrm, TB_ALIGN_32 },
1027 { X86::VPADDQYrr, X86::VPADDQYrm, TB_ALIGN_32 },
1028 { X86::VPADDSBYrr, X86::VPADDSBYrm, TB_ALIGN_32 },
1029 { X86::VPADDSWYrr, X86::VPADDSWYrm, TB_ALIGN_32 },
1030 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, TB_ALIGN_32 },
1031 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, TB_ALIGN_32 },
1032 { X86::VPADDWYrr, X86::VPADDWYrm, TB_ALIGN_32 },
1033 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, TB_ALIGN_32 },
1034 { X86::VPANDNYrr, X86::VPANDNYrm, TB_ALIGN_32 },
1035 { X86::VPANDYrr, X86::VPANDYrm, TB_ALIGN_32 },
1036 { X86::VPAVGBYrr, X86::VPAVGBYrm, TB_ALIGN_32 },
1037 { X86::VPAVGWYrr, X86::VPAVGWYrm, TB_ALIGN_32 },
1038 { X86::VPBLENDDrri, X86::VPBLENDDrmi, TB_ALIGN_32 },
1039 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, TB_ALIGN_32 },
1040 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, TB_ALIGN_32 },
1041 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, TB_ALIGN_32 },
1042 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, TB_ALIGN_32 },
1043 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, TB_ALIGN_32 },
1044 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, TB_ALIGN_32 },
1045 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, TB_ALIGN_32 },
1046 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, TB_ALIGN_32 },
1047 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, TB_ALIGN_32 },
1048 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, TB_ALIGN_32 },
1049 { X86::VPERM2I128rr, X86::VPERM2I128rm, TB_ALIGN_32 },
Craig Topper40385c82012-01-19 08:50:38 +0000[diff] [blame]1050 { X86::VPERMDYrr, X86::VPERMDYrm, TB_ALIGN_32 },
Elena Demikhovsky73c504a2012-04-15 11:18:59 +0000[diff] [blame]1051 { X86::VPERMPDYri, X86::VPERMPDYmi, TB_ALIGN_32 },
Craig Topper40385c82012-01-19 08:50:38 +0000[diff] [blame]1052 { X86::VPERMPSYrr, X86::VPERMPSYrm, TB_ALIGN_32 },
Elena Demikhovsky73c504a2012-04-15 11:18:59 +0000[diff] [blame]1053 { X86::VPERMQYri, X86::VPERMQYmi, TB_ALIGN_32 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]1054 { X86::VPHADDDYrr, X86::VPHADDDYrm, TB_ALIGN_32 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1055 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, TB_ALIGN_32 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]1056 { X86::VPHADDWYrr, X86::VPHADDWYrm, TB_ALIGN_32 },
1057 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, TB_ALIGN_32 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1058 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, TB_ALIGN_32 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000[diff] [blame]1059 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, TB_ALIGN_32 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1060 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, TB_ALIGN_32 },
1061 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, TB_ALIGN_32 },
1062 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, TB_ALIGN_32 },
1063 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, TB_ALIGN_32 },
1064 { X86::VPMINSWYrr, X86::VPMINSWYrm, TB_ALIGN_32 },
1065 { X86::VPMINUBYrr, X86::VPMINUBYrm, TB_ALIGN_32 },
1066 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, TB_ALIGN_32 },
1067 { X86::VPMULDQYrr, X86::VPMULDQYrm, TB_ALIGN_32 },
1068 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, TB_ALIGN_32 },
1069 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, TB_ALIGN_32 },
1070 { X86::VPMULHWYrr, X86::VPMULHWYrm, TB_ALIGN_32 },
1071 { X86::VPMULLDYrr, X86::VPMULLDYrm, TB_ALIGN_32 },
1072 { X86::VPMULLWYrr, X86::VPMULLWYrm, TB_ALIGN_32 },
1073 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, TB_ALIGN_32 },
1074 { X86::VPORYrr, X86::VPORYrm, TB_ALIGN_32 },
1075 { X86::VPSADBWYrr, X86::VPSADBWYrm, TB_ALIGN_32 },
Craig Topper969ba282012-01-25 06:43:11 +0000[diff] [blame]1076 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, TB_ALIGN_32 },
1077 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, TB_ALIGN_32 },
1078 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, TB_ALIGN_32 },
1079 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1080 { X86::VPSLLDYrr, X86::VPSLLDYrm, TB_ALIGN_16 },
1081 { X86::VPSLLQYrr, X86::VPSLLQYrm, TB_ALIGN_16 },
1082 { X86::VPSLLWYrr, X86::VPSLLWYrm, TB_ALIGN_16 },
1083 { X86::VPSLLVDrr, X86::VPSLLVDrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1084 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1085 { X86::VPSLLVQrr, X86::VPSLLVQrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1086 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1087 { X86::VPSRADYrr, X86::VPSRADYrm, TB_ALIGN_16 },
1088 { X86::VPSRAWYrr, X86::VPSRAWYrm, TB_ALIGN_16 },
1089 { X86::VPSRAVDrr, X86::VPSRAVDrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1090 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1091 { X86::VPSRLDYrr, X86::VPSRLDYrm, TB_ALIGN_16 },
1092 { X86::VPSRLQYrr, X86::VPSRLQYrm, TB_ALIGN_16 },
1093 { X86::VPSRLWYrr, X86::VPSRLWYrm, TB_ALIGN_16 },
1094 { X86::VPSRLVDrr, X86::VPSRLVDrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1095 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1096 { X86::VPSRLVQrr, X86::VPSRLVQrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1097 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, TB_ALIGN_32 },
1098 { X86::VPSUBBYrr, X86::VPSUBBYrm, TB_ALIGN_32 },
1099 { X86::VPSUBDYrr, X86::VPSUBDYrm, TB_ALIGN_32 },
1100 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, TB_ALIGN_32 },
1101 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, TB_ALIGN_32 },
1102 { X86::VPSUBWYrr, X86::VPSUBWYrm, TB_ALIGN_32 },
1103 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, TB_ALIGN_32 },
1104 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000[diff] [blame]1105 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000[diff] [blame]1106 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, TB_ALIGN_32 },
1107 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, TB_ALIGN_32 },
1108 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, TB_ALIGN_32 },
1109 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, TB_ALIGN_32 },
1110 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, TB_ALIGN_32 },
1111 { X86::VPXORYrr, X86::VPXORYrm, TB_ALIGN_32 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]1112 // FIXME: add AVX 256-bit foldable instructions
Craig Topperdfb1e4b2012-08-31 23:10:34 +0000[diff] [blame]1113
1114 // FMA4 foldable patterns
1115 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, TB_ALIGN_16 },
1116 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, TB_ALIGN_16 },
1117 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1118 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1119 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1120 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
1121 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1122 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1123 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1124 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
1125 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, TB_ALIGN_16 },
1126 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, TB_ALIGN_16 },
1127 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1128 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1129 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1130 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
1131 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1132 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1133 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1134 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1135 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1136 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1137 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1138 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1139 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1140 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1141 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1142 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]1143 };
1144
1145 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]1146 unsigned RegOp = OpTbl2[i].RegOp;
1147 unsigned MemOp = OpTbl2[i].MemOp;
1148 unsigned Flags = OpTbl2[i].Flags;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]1149 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1150 RegOp, MemOp,
1151 // Index 2, folded load
1152 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]1153 }
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1154
1155 static const X86OpTblEntry OpTbl3[] = {
1156 // FMA foldable instructions
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1157 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1158 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1159 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1160 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1161 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1162 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
Elena Demikhovsky1503aba2012-08-01 12:06:00 +0000[diff] [blame]1163 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, 0 },
1164 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, 0 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1165
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1166 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1167 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1168 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1169 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1170 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1171 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1172 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1173 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1174 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1175 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1176 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1177 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1178
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1179 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1180 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1181 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1182 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1183 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1184 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
Elena Demikhovsky1503aba2012-08-01 12:06:00 +0000[diff] [blame]1185 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, 0 },
1186 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, 0 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1187
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1188 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1189 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1190 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1191 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1192 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1193 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1194 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1195 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1196 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1197 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1198 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1199 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1200
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1201 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1202 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1203 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1204 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1205 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1206 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
Elena Demikhovsky1503aba2012-08-01 12:06:00 +0000[diff] [blame]1207 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, 0 },
1208 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, 0 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1209
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1210 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1211 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1212 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1213 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1214 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1215 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1216 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1217 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1218 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1219 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1220 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1221 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1222
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1223 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1224 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1225 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1226 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1227 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1228 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
Elena Demikhovsky1503aba2012-08-01 12:06:00 +0000[diff] [blame]1229 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, 0 },
1230 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, 0 },
Craig Topper78fc72d2012-06-01 05:48:39 +0000[diff] [blame]1231
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1232 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1233 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1234 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1235 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1236 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1237 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1238 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1239 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1240 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1241 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1242 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1243 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
Craig Topperfc5ab242012-06-04 07:08:21 +0000[diff] [blame]1244
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1245 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1246 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1247 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1248 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1249 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1250 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1251 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1252 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1253 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1254 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1255 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1256 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
Craig Topperfc5ab242012-06-04 07:08:21 +0000[diff] [blame]1257
Craig Toppercaea5e22012-06-04 07:46:16 +0000[diff] [blame]1258 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1259 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1260 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1261 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1262 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1263 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1264 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1265 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1266 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1267 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1268 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1269 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
Craig Topperdfb1e4b2012-08-31 23:10:34 +0000[diff] [blame]1270
1271 // FMA4 foldable patterns
1272 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, TB_ALIGN_16 },
1273 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, TB_ALIGN_16 },
1274 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1275 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1276 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1277 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
1278 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1279 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1280 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1281 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
1282 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, TB_ALIGN_16 },
1283 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, TB_ALIGN_16 },
1284 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1285 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1286 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1287 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
1288 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1289 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1290 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1291 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1292 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1293 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1294 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1295 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1296 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1297 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1298 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1299 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +0000[diff] [blame]1300 };
1301
1302 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1303 unsigned RegOp = OpTbl3[i].RegOp;
1304 unsigned MemOp = OpTbl3[i].MemOp;
1305 unsigned Flags = OpTbl3[i].Flags;
1306 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1307 RegOp, MemOp,
1308 // Index 3, folded load
1309 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1310 }
1311
Chris Lattner72614082002-10-25 22:55:53 +0000[diff] [blame]1312}
1313
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]1314void
1315X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1316 MemOp2RegOpTableType &M2RTable,
1317 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1318 if ((Flags & TB_NO_FORWARD) == 0) {
1319 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1320 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1321 }
1322 if ((Flags & TB_NO_REVERSE) == 0) {
1323 assert(!M2RTable.count(MemOp) &&
1324 "Duplicated entries in unfolding maps?");
1325 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1326 }
1327}
1328
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1329bool
Evan Cheng7da9ecf2010-01-13 00:30:23 +0000[diff] [blame]1330X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1331 unsigned &SrcReg, unsigned &DstReg,
1332 unsigned &SubIdx) const {
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1333 switch (MI.getOpcode()) {
1334 default: break;
1335 case X86::MOVSX16rr8:
1336 case X86::MOVZX16rr8:
1337 case X86::MOVSX32rr8:
1338 case X86::MOVZX32rr8:
1339 case X86::MOVSX64rr8:
1340 case X86::MOVZX64rr8:
Evan Cheng57d1d932010-01-13 08:01:32 +0000[diff] [blame]1341 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1342 // It's not always legal to reference the low 8-bit of the larger
1343 // register in 32-bit mode.
1344 return false;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1345 case X86::MOVSX32rr16:
1346 case X86::MOVZX32rr16:
1347 case X86::MOVSX64rr16:
1348 case X86::MOVZX64rr16:
1349 case X86::MOVSX64rr32:
1350 case X86::MOVZX64rr32: {
1351 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1352 // Be conservative.
1353 return false;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1354 SrcReg = MI.getOperand(1).getReg();
1355 DstReg = MI.getOperand(0).getReg();
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1356 switch (MI.getOpcode()) {
Craig Topper195f1b82012-08-21 08:16:16 +0000[diff] [blame]1357 default: llvm_unreachable("Unreachable!");
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1358 case X86::MOVSX16rr8:
1359 case X86::MOVZX16rr8:
1360 case X86::MOVSX32rr8:
1361 case X86::MOVZX32rr8:
1362 case X86::MOVSX64rr8:
1363 case X86::MOVZX64rr8:
Jakob Stoklund Olesen22c0e972010-05-25 17:04:16 +0000[diff] [blame]1364 SubIdx = X86::sub_8bit;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1365 break;
1366 case X86::MOVSX32rr16:
1367 case X86::MOVZX32rr16:
1368 case X86::MOVSX64rr16:
1369 case X86::MOVZX64rr16:
Jakob Stoklund Olesen22c0e972010-05-25 17:04:16 +0000[diff] [blame]1370 SubIdx = X86::sub_16bit;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1371 break;
1372 case X86::MOVSX64rr32:
1373 case X86::MOVZX64rr32:
Jakob Stoklund Olesen22c0e972010-05-25 17:04:16 +0000[diff] [blame]1374 SubIdx = X86::sub_32bit;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1375 break;
1376 }
Evan Cheng7da9ecf2010-01-13 00:30:23 +0000[diff] [blame]1377 return true;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1378 }
1379 }
Evan Cheng7da9ecf2010-01-13 00:30:23 +0000[diff] [blame]1380 return false;
Evan Chenga5a81d72010-01-12 00:09:37 +0000[diff] [blame]1381}
1382
David Greeneb87bc952009-11-12 20:55:29 +0000[diff] [blame]1383/// isFrameOperand - Return true and the FrameIndex if the specified
1384/// operand and follow operands form a reference to the stack frame.
1385bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1386 int &FrameIndex) const {
1387 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1388 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1389 MI->getOperand(Op+1).getImm() == 1 &&
1390 MI->getOperand(Op+2).getReg() == 0 &&
1391 MI->getOperand(Op+3).getImm() == 0) {
1392 FrameIndex = MI->getOperand(Op).getIndex();
1393 return true;
1394 }
1395 return false;
1396}
1397
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1398static bool isFrameLoadOpcode(int Opcode) {
1399 switch (Opcode) {
David Blaikie4d6ccb52012-01-20 21:51:11 +0000[diff] [blame]1400 default:
1401 return false;
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1402 case X86::MOV8rm:
1403 case X86::MOV16rm:
1404 case X86::MOV32rm:
Evan Cheng25ab6902006-09-08 06:48:29 +0000[diff] [blame]1405 case X86::MOV64rm:
Dale Johannesene377d4d2007-07-04 21:07:47 +0000[diff] [blame]1406 case X86::LD_Fp64m:
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1407 case X86::MOVSSrm:
1408 case X86::MOVSDrm:
Chris Lattner993c8972006-04-18 16:44:51 +0000[diff] [blame]1409 case X86::MOVAPSrm:
1410 case X86::MOVAPDrm:
Dan Gohman54462742009-01-09 02:40:34 +0000[diff] [blame]1411 case X86::MOVDQArm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]1412 case X86::VMOVSSrm:
1413 case X86::VMOVSDrm:
1414 case X86::VMOVAPSrm:
1415 case X86::VMOVAPDrm:
1416 case X86::VMOVDQArm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]1417 case X86::VMOVAPSYrm:
1418 case X86::VMOVAPDYrm:
1419 case X86::VMOVDQAYrm:
Bill Wendling823efee2007-04-03 06:00:37 +0000[diff] [blame]1420 case X86::MMX_MOVD64rm:
1421 case X86::MMX_MOVQ64rm:
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1422 return true;
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1423 }
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1424}
1425
1426static bool isFrameStoreOpcode(int Opcode) {
1427 switch (Opcode) {
1428 default: break;
1429 case X86::MOV8mr:
1430 case X86::MOV16mr:
1431 case X86::MOV32mr:
1432 case X86::MOV64mr:
1433 case X86::ST_FpP64m:
1434 case X86::MOVSSmr:
1435 case X86::MOVSDmr:
1436 case X86::MOVAPSmr:
1437 case X86::MOVAPDmr:
1438 case X86::MOVDQAmr:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]1439 case X86::VMOVSSmr:
1440 case X86::VMOVSDmr:
1441 case X86::VMOVAPSmr:
1442 case X86::VMOVAPDmr:
1443 case X86::VMOVDQAmr:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]1444 case X86::VMOVAPSYmr:
1445 case X86::VMOVAPDYmr:
1446 case X86::VMOVDQAYmr:
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1447 case X86::MMX_MOVD64mr:
1448 case X86::MMX_MOVQ64mr:
1449 case X86::MMX_MOVNTQmr:
1450 return true;
1451 }
1452 return false;
1453}
1454
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1455unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1456 int &FrameIndex) const {
1457 if (isFrameLoadOpcode(MI->getOpcode()))
Jakob Stoklund Olesen81c7b192010-07-27 04:17:01 +0000[diff] [blame]1458 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1459 return MI->getOperand(0).getReg();
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1460 return 0;
1461}
1462
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1463unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1464 int &FrameIndex) const {
1465 if (isFrameLoadOpcode(MI->getOpcode())) {
1466 unsigned Reg;
1467 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1468 return Reg;
David Greeneb87bc952009-11-12 20:55:29 +0000[diff] [blame]1469 // Check for post-frame index elimination operations
David Greene29dbf502009-12-04 22:38:46 +0000[diff] [blame]1470 const MachineMemOperand *Dummy;
1471 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1472 }
1473 return 0;
1474}
1475
Dan Gohmancbad42c2008-11-18 19:49:32 +0000[diff] [blame]1476unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1477 int &FrameIndex) const {
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1478 if (isFrameStoreOpcode(MI->getOpcode()))
Jakob Stoklund Olesen81c7b192010-07-27 04:17:01 +0000[diff] [blame]1479 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1480 isFrameOperand(MI, 0, FrameIndex))
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]1481 return MI->getOperand(X86::AddrNumOperands).getReg();
David Greenedda39782009-11-13 00:29:53 +0000[diff] [blame]1482 return 0;
1483}
1484
1485unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1486 int &FrameIndex) const {
1487 if (isFrameStoreOpcode(MI->getOpcode())) {
1488 unsigned Reg;
1489 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1490 return Reg;
David Greeneb87bc952009-11-12 20:55:29 +0000[diff] [blame]1491 // Check for post-frame index elimination operations
David Greene29dbf502009-12-04 22:38:46 +0000[diff] [blame]1492 const MachineMemOperand *Dummy;
1493 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
Chris Lattner40839602006-02-02 20:12:32 +0000[diff] [blame]1494 }
1495 return 0;
1496}
1497
Evan Chenge3d8dbf2008-03-27 01:45:11 +0000[diff] [blame]1498/// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1499/// X86::MOVPC32r.
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]1500static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
Jakob Stoklund Olesen130e6032012-08-08 00:40:47 +0000[diff] [blame]1501 // Don't waste compile time scanning use-def chains of physregs.
1502 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1503 return false;
Evan Chenge3d8dbf2008-03-27 01:45:11 +0000[diff] [blame]1504 bool isPICBase = false;
1505 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1506 E = MRI.def_end(); I != E; ++I) {
1507 MachineInstr *DefMI = I.getOperand().getParent();
1508 if (DefMI->getOpcode() != X86::MOVPC32r)
1509 return false;
1510 assert(!isPICBase && "More than one PIC base?");
1511 isPICBase = true;
1512 }
1513 return isPICBase;
1514}
Evan Cheng9d15abe2008-03-31 07:54:19 +0000[diff] [blame]1515
Bill Wendling9f8fea32008-05-12 20:54:26 +0000[diff] [blame]1516bool
Dan Gohman3731bc02009-10-10 00:34:18 +0000[diff] [blame]1517X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1518 AliasAnalysis *AA) const {
Dan Gohmanc101e952007-06-14 20:50:44 +0000[diff] [blame]1519 switch (MI->getOpcode()) {
1520 default: break;
Craig Topper630e33a2012-08-21 08:17:07 +0000[diff] [blame]1521 case X86::MOV8rm:
1522 case X86::MOV16rm:
1523 case X86::MOV32rm:
1524 case X86::MOV64rm:
1525 case X86::LD_Fp64m:
1526 case X86::MOVSSrm:
1527 case X86::MOVSDrm:
1528 case X86::MOVAPSrm:
1529 case X86::MOVUPSrm:
1530 case X86::MOVAPDrm:
1531 case X86::MOVDQArm:
1532 case X86::VMOVSSrm:
1533 case X86::VMOVSDrm:
1534 case X86::VMOVAPSrm:
1535 case X86::VMOVUPSrm:
1536 case X86::VMOVAPDrm:
1537 case X86::VMOVDQArm:
1538 case X86::VMOVAPSYrm:
1539 case X86::VMOVUPSYrm:
1540 case X86::VMOVAPDYrm:
1541 case X86::VMOVDQAYrm:
1542 case X86::MMX_MOVD64rm:
1543 case X86::MMX_MOVQ64rm:
1544 case X86::FsVMOVAPSrm:
1545 case X86::FsVMOVAPDrm:
1546 case X86::FsMOVAPSrm:
1547 case X86::FsMOVAPDrm: {
1548 // Loads from constant pools are trivially rematerializable.
1549 if (MI->getOperand(1).isReg() &&
1550 MI->getOperand(2).isImm() &&
1551 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1552 MI->isInvariantLoad(AA)) {
1553 unsigned BaseReg = MI->getOperand(1).getReg();
1554 if (BaseReg == 0 || BaseReg == X86::RIP)
1555 return true;
1556 // Allow re-materialization of PIC load.
1557 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1558 return false;
1559 const MachineFunction &MF = *MI->getParent()->getParent();
1560 const MachineRegisterInfo &MRI = MF.getRegInfo();
1561 return regIsPICBase(BaseReg, MRI);
Evan Chengd8850a52008-02-22 09:25:47 +0000[diff] [blame]1562 }
Craig Topper630e33a2012-08-21 08:17:07 +0000[diff] [blame]1563 return false;
1564 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1565
Craig Topper630e33a2012-08-21 08:17:07 +0000[diff] [blame]1566 case X86::LEA32r:
1567 case X86::LEA64r: {
1568 if (MI->getOperand(2).isImm() &&
1569 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1570 !MI->getOperand(4).isReg()) {
1571 // lea fi#, lea GV, etc. are all rematerializable.
1572 if (!MI->getOperand(1).isReg())
1573 return true;
1574 unsigned BaseReg = MI->getOperand(1).getReg();
1575 if (BaseReg == 0)
1576 return true;
1577 // Allow re-materialization of lea PICBase + x.
1578 const MachineFunction &MF = *MI->getParent()->getParent();
1579 const MachineRegisterInfo &MRI = MF.getRegInfo();
1580 return regIsPICBase(BaseReg, MRI);
1581 }
1582 return false;
1583 }
Dan Gohmanc101e952007-06-14 20:50:44 +0000[diff] [blame]1584 }
Evan Chenge771ebd2008-03-27 01:41:09 +0000[diff] [blame]1585
Dan Gohmand45eddd2007-06-26 00:48:07 +0000[diff] [blame]1586 // All other instructions marked M_REMATERIALIZABLE are always trivially
1587 // rematerializable.
1588 return true;
Dan Gohmanc101e952007-06-14 20:50:44 +0000[diff] [blame]1589}
1590
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1591/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1592/// would clobber the EFLAGS condition register. Note the result may be
1593/// conservative. If it cannot definitely determine the safety after visiting
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1594/// a few instructions in each direction it assumes it's not safe.
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1595static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1596 MachineBasicBlock::iterator I) {
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1597 MachineBasicBlock::iterator E = MBB.end();
1598
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1599 // For compile time consideration, if we are not able to determine the
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1600 // safety after visiting 4 instructions in each direction, we will assume
1601 // it's not safe.
1602 MachineBasicBlock::iterator Iter = I;
Jakob Stoklund Olesenb8e052e2011-09-02 23:52:52 +0000[diff] [blame]1603 for (unsigned i = 0; Iter != E && i < 4; ++i) {
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1604 bool SeenDef = false;
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1605 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1606 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen450b3852012-02-09 00:17:22 +0000[diff] [blame]1607 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1608 SeenDef = true;
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]1609 if (!MO.isReg())
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1610 continue;
1611 if (MO.getReg() == X86::EFLAGS) {
1612 if (MO.isUse())
1613 return false;
1614 SeenDef = true;
1615 }
1616 }
1617
1618 if (SeenDef)
1619 // This instruction defines EFLAGS, no need to look any further.
1620 return true;
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1621 ++Iter;
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1622 // Skip over DBG_VALUE.
1623 while (Iter != E && Iter->isDebugValue())
1624 ++Iter;
Jakob Stoklund Olesenb8e052e2011-09-02 23:52:52 +0000[diff] [blame]1625 }
Dan Gohman3afda6e2008-10-21 03:24:31 +0000[diff] [blame]1626
Jakob Stoklund Olesenb8e052e2011-09-02 23:52:52 +0000[diff] [blame]1627 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1628 // live in.
1629 if (Iter == E) {
1630 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1631 SE = MBB.succ_end(); SI != SE; ++SI)
1632 if ((*SI)->isLiveIn(X86::EFLAGS))
1633 return false;
1634 return true;
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1635 }
1636
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1637 MachineBasicBlock::iterator B = MBB.begin();
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1638 Iter = I;
1639 for (unsigned i = 0; i < 4; ++i) {
1640 // If we make it to the beginning of the block, it's safe to clobber
1641 // EFLAGS iff EFLAGS is not live-in.
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1642 if (Iter == B)
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1643 return !MBB.isLiveIn(X86::EFLAGS);
1644
1645 --Iter;
Evan Cheng8d1f0dd2010-03-23 20:35:45 +0000[diff] [blame]1646 // Skip over DBG_VALUE.
1647 while (Iter != B && Iter->isDebugValue())
1648 --Iter;
1649
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1650 bool SawKill = false;
1651 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1652 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen450b3852012-02-09 00:17:22 +0000[diff] [blame]1653 // A register mask may clobber EFLAGS, but we should still look for a
1654 // live EFLAGS def.
1655 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1656 SawKill = true;
Dan Gohman1b1764b2009-10-14 00:08:59 +0000[diff] [blame]1657 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1658 if (MO.isDef()) return MO.isDead();
1659 if (MO.isKill()) SawKill = true;
1660 }
1661 }
1662
1663 if (SawKill)
1664 // This instruction kills EFLAGS and doesn't redefine it, so
1665 // there's no need to look further.
Dan Gohman3afda6e2008-10-21 03:24:31 +0000[diff] [blame]1666 return true;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1667 }
1668
1669 // Conservative answer.
1670 return false;
1671}
1672
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1673void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1674 MachineBasicBlock::iterator I,
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1675 unsigned DestReg, unsigned SubIdx,
Evan Chengd57cdd52009-11-14 02:55:43 +0000[diff] [blame]1676 const MachineInstr *Orig,
Jakob Stoklund Olesen9edf7de2010-06-02 22:47:25 +0000[diff] [blame]1677 const TargetRegisterInfo &TRI) const {
Dan Gohman0d881042010-05-07 01:28:10 +0000[diff] [blame]1678 DebugLoc DL = Orig->getDebugLoc();
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1679
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1680 // MOV32r0 etc. are implemented with xor which clobbers condition code.
1681 // Re-materialize them as movri instructions to avoid side effects.
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1682 bool Clone = true;
1683 unsigned Opc = Orig->getOpcode();
1684 switch (Opc) {
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1685 default: break;
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1686 case X86::MOV8r0:
Dan Gohmanf1b4d262010-01-12 04:42:54 +0000[diff] [blame]1687 case X86::MOV16r0:
1688 case X86::MOV32r0:
1689 case X86::MOV64r0: {
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1690 if (!isSafeToClobberEFLAGS(MBB, I)) {
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1691 switch (Opc) {
Craig Topper195f1b82012-08-21 08:16:16 +0000[diff] [blame]1692 default: llvm_unreachable("Unreachable!");
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1693 case X86::MOV8r0: Opc = X86::MOV8ri; break;
Dan Gohmanf1b4d262010-01-12 04:42:54 +0000[diff] [blame]1694 case X86::MOV16r0: Opc = X86::MOV16ri; break;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1695 case X86::MOV32r0: Opc = X86::MOV32ri; break;
Dan Gohman6fe0df22010-02-26 16:49:27 +0000[diff] [blame]1696 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1697 }
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1698 Clone = false;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1699 }
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1700 break;
Evan Cheng9ef4ca22008-06-24 07:10:51 +0000[diff] [blame]1701 }
1702 }
1703
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1704 if (Clone) {
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]1705 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1706 MBB.insert(I, MI);
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1707 } else {
Jakob Stoklund Olesen9edf7de2010-06-02 22:47:25 +0000[diff] [blame]1708 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0);
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1709 }
Evan Cheng03eb3882008-04-16 23:44:44 +0000[diff] [blame]1710
Evan Cheng37844532009-07-16 09:20:10 +0000[diff] [blame]1711 MachineInstr *NewMI = prior(I);
Jakob Stoklund Olesen9edf7de2010-06-02 22:47:25 +0000[diff] [blame]1712 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
Evan Chengca1267c2008-03-31 20:40:39 +0000[diff] [blame]1713}
1714
Evan Cheng3f411c72007-10-05 08:04:01 +0000[diff] [blame]1715/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1716/// is not marked dead.
1717static bool hasLiveCondCodeDef(MachineInstr *MI) {
Evan Cheng3f411c72007-10-05 08:04:01 +0000[diff] [blame]1718 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1719 MachineOperand &MO = MI->getOperand(i);
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]1720 if (MO.isReg() && MO.isDef() &&
Evan Cheng3f411c72007-10-05 08:04:01 +0000[diff] [blame]1721 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1722 return true;
1723 }
1724 }
1725 return false;
1726}
1727
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1728/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1729/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1730/// to a 32-bit superregister and then truncating back down to a 16-bit
1731/// subregister.
1732MachineInstr *
1733X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1734 MachineFunction::iterator &MFI,
1735 MachineBasicBlock::iterator &MBBI,
1736 LiveVariables *LV) const {
1737 MachineInstr *MI = MBBI;
1738 unsigned Dest = MI->getOperand(0).getReg();
1739 unsigned Src = MI->getOperand(1).getReg();
1740 bool isDead = MI->getOperand(0).isDead();
1741 bool isKill = MI->getOperand(1).isKill();
1742
1743 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1744 ? X86::LEA64_32r : X86::LEA32r;
1745 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1746 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1747 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1748
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1749 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1750 // well be shifting and then extracting the lower 16-bits.
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1751 // This has the potential to cause partial register stall. e.g.
Evan Cheng04ab19c2009-12-12 18:55:26 +0000[diff] [blame]1752 // movw (%rbp,%rcx,2), %dx
1753 // leal -65(%rdx), %esi
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1754 // But testing has shown this *does* help performance in 64-bit mode (at
1755 // least on modern x86 machines).
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1756 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1757 MachineInstr *InsMI =
Jakob Stoklund Olesen5c00e072010-07-08 16:40:15 +0000[diff] [blame]1758 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1759 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1760 .addReg(Src, getKillRegState(isKill));
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1761
1762 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1763 get(Opc), leaOutReg);
1764 switch (MIOpc) {
Craig Topper195f1b82012-08-21 08:16:16 +0000[diff] [blame]1765 default: llvm_unreachable("Unreachable!");
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1766 case X86::SHL16ri: {
1767 unsigned ShAmt = MI->getOperand(2).getImm();
1768 MIB.addReg(0).addImm(1 << ShAmt)
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1769 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1770 break;
1771 }
1772 case X86::INC16r:
1773 case X86::INC64_16r:
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1774 addRegOffset(MIB, leaInReg, true, 1);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1775 break;
1776 case X86::DEC16r:
1777 case X86::DEC64_16r:
Chris Lattner599b5312010-07-08 23:46:44 +0000[diff] [blame]1778 addRegOffset(MIB, leaInReg, true, -1);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1779 break;
1780 case X86::ADD16ri:
1781 case X86::ADD16ri8:
Chris Lattner15df55d2010-10-08 03:57:25 +0000[diff] [blame]1782 case X86::ADD16ri_DB:
1783 case X86::ADD16ri8_DB:
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1784 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1785 break;
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]1786 case X86::ADD16rr:
1787 case X86::ADD16rr_DB: {
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1788 unsigned Src2 = MI->getOperand(2).getReg();
1789 bool isKill2 = MI->getOperand(2).isKill();
1790 unsigned leaInReg2 = 0;
1791 MachineInstr *InsMI2 = 0;
1792 if (Src == Src2) {
1793 // ADD16rr %reg1028<kill>, %reg1028
1794 // just a single insert_subreg.
1795 addRegReg(MIB, leaInReg, true, leaInReg, false);
1796 } else {
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1797 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1798 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1799 // well be shifting and then extracting the lower 16-bits.
Evan Chengddfd1372011-12-14 02:11:42 +0000[diff] [blame]1800 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1801 InsMI2 =
Evan Chengddfd1372011-12-14 02:11:42 +0000[diff] [blame]1802 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
Jakob Stoklund Olesen5c00e072010-07-08 16:40:15 +0000[diff] [blame]1803 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1804 .addReg(Src2, getKillRegState(isKill2));
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1805 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1806 }
1807 if (LV && isKill2 && InsMI2)
1808 LV->replaceKillInstruction(Src2, MI, InsMI2);
1809 break;
1810 }
1811 }
1812
1813 MachineInstr *NewMI = MIB;
1814 MachineInstr *ExtMI =
Jakob Stoklund Olesen0bc25f42010-07-08 16:40:22 +0000[diff] [blame]1815 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1816 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Jakob Stoklund Olesen0bc25f42010-07-08 16:40:22 +0000[diff] [blame]1817 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1818
1819 if (LV) {
1820 // Update live variables
1821 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1822 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1823 if (isKill)
1824 LV->replaceKillInstruction(Src, MI, InsMI);
1825 if (isDead)
1826 LV->replaceKillInstruction(Dest, MI, ExtMI);
1827 }
1828
1829 return ExtMI;
1830}
1831
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]1832/// convertToThreeAddress - This method must be implemented by targets that
1833/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1834/// may be able to convert a two-address instruction into a true
1835/// three-address instruction on demand. This allows the X86 target (for
1836/// example) to convert ADD and SHL instructions into LEA instructions if they
1837/// would require register copies due to two-addressness.
1838///
1839/// This method returns a null pointer if the transformation cannot be
1840/// performed, otherwise it returns the new instruction.
1841///
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1842MachineInstr *
1843X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1844 MachineBasicBlock::iterator &MBBI,
Owen Andersonf660c172008-07-02 23:41:07 +0000[diff] [blame]1845 LiveVariables *LV) const {
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1846 MachineInstr *MI = MBBI;
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]1847 MachineFunction &MF = *MI->getParent()->getParent();
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]1848 // All instructions input are two-addr instructions. Get the known operands.
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1849 const MachineOperand &Dest = MI->getOperand(0);
1850 const MachineOperand &Src = MI->getOperand(1);
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]1851
Evan Cheng6ce7dc22006-11-15 20:58:11 +0000[diff] [blame]1852 MachineInstr *NewMI = NULL;
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1853 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1854 // we have better subtarget support, enable the 16-bit LEA generation here.
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1855 // 16-bit LEA is also slow on Core2.
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1856 bool DisableLEA16 = true;
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1857 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
Evan Cheng258ff672006-12-01 21:52:41 +0000[diff] [blame]1858
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1859 unsigned MIOpc = MI->getOpcode();
1860 switch (MIOpc) {
Evan Chengccba76b2006-05-30 20:26:50 +0000[diff] [blame]1861 case X86::SHUFPSrri: {
1862 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1863 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]1864
Evan Chengaa3c1412006-05-30 21:45:53 +0000[diff] [blame]1865 unsigned B = MI->getOperand(1).getReg();
1866 unsigned C = MI->getOperand(2).getReg();
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1867 if (B != C) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1868 unsigned M = MI->getOperand(3).getImm();
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1869 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1870 .addOperand(Dest).addOperand(Src).addImm(M);
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1871 break;
1872 }
Craig Topper05189702012-01-13 09:21:41 +0000[diff] [blame]1873 case X86::SHUFPDrri: {
1874 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1875 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1876
1877 unsigned B = MI->getOperand(1).getReg();
1878 unsigned C = MI->getOperand(2).getReg();
1879 if (B != C) return 0;
Craig Topper05189702012-01-13 09:21:41 +0000[diff] [blame]1880 unsigned M = MI->getOperand(3).getImm();
1881
1882 // Convert to PSHUFD mask.
1883 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1884
1885 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1886 .addOperand(Dest).addOperand(Src).addImm(M);
Craig Topper05189702012-01-13 09:21:41 +0000[diff] [blame]1887 break;
1888 }
Chris Lattner995f5502007-03-28 18:12:31 +0000[diff] [blame]1889 case X86::SHL64ri: {
Evan Cheng24f2ea32007-09-14 21:48:26 +0000[diff] [blame]1890 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattner995f5502007-03-28 18:12:31 +0000[diff] [blame]1891 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1892 // the flags produced by a shift yet, so this is safe.
Chris Lattner995f5502007-03-28 18:12:31 +0000[diff] [blame]1893 unsigned ShAmt = MI->getOperand(2).getImm();
1894 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1895
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1896 // LEA can't handle RSP.
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1897 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1898 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1899 &X86::GR64_NOSPRegClass))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1900 return 0;
1901
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1902 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1903 .addOperand(Dest)
1904 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattner995f5502007-03-28 18:12:31 +0000[diff] [blame]1905 break;
1906 }
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1907 case X86::SHL32ri: {
Evan Cheng24f2ea32007-09-14 21:48:26 +0000[diff] [blame]1908 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1909 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1910 // the flags produced by a shift yet, so this is safe.
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1911 unsigned ShAmt = MI->getOperand(2).getImm();
1912 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1913
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1914 // LEA can't handle ESP.
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1915 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1916 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1917 &X86::GR32_NOSPRegClass))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1918 return 0;
1919
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1920 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]1921 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1922 .addOperand(Dest)
1923 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1924 break;
1925 }
1926 case X86::SHL16ri: {
Evan Cheng24f2ea32007-09-14 21:48:26 +0000[diff] [blame]1927 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Evan Cheng61d9c862007-09-06 00:14:41 +0000[diff] [blame]1928 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1929 // the flags produced by a shift yet, so this is safe.
Evan Cheng61d9c862007-09-06 00:14:41 +0000[diff] [blame]1930 unsigned ShAmt = MI->getOperand(2).getImm();
1931 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]1932
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1933 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1934 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1935 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1936 .addOperand(Dest)
1937 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
Chris Lattnera16b7cb2007-03-20 06:08:29 +0000[diff] [blame]1938 break;
Evan Chengccba76b2006-05-30 20:26:50 +0000[diff] [blame]1939 }
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1940 default: {
1941 // The following opcodes also sets the condition code register(s). Only
1942 // convert them to equivalent lea if the condition code register def's
1943 // are dead!
1944 if (hasLiveCondCodeDef(MI))
1945 return 0;
Evan Chengccba76b2006-05-30 20:26:50 +0000[diff] [blame]1946
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1947 switch (MIOpc) {
1948 default: return 0;
1949 case X86::INC64r:
Dan Gohmancca29832009-01-06 23:34:46 +0000[diff] [blame]1950 case X86::INC32r:
1951 case X86::INC64_32r: {
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1952 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Evan Chengb76143c2007-10-09 07:14:53 +0000[diff] [blame]1953 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1954 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]1955 const TargetRegisterClass *RC = MIOpc == X86::INC64r ?
1956 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1957 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1958
1959 // LEA can't handle RSP.
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1960 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1961 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1962 return 0;
1963
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1964 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1965 .addOperand(Dest).addOperand(Src), 1);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1966 break;
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]1967 }
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1968 case X86::INC16r:
1969 case X86::INC64_16r:
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1970 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1971 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1972 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1973 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1974 .addOperand(Dest).addOperand(Src), 1);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1975 break;
1976 case X86::DEC64r:
Dan Gohmancca29832009-01-06 23:34:46 +0000[diff] [blame]1977 case X86::DEC32r:
1978 case X86::DEC64_32r: {
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1979 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Evan Chengb76143c2007-10-09 07:14:53 +0000[diff] [blame]1980 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1981 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]1982 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ?
1983 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1984 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1985 // LEA can't handle RSP.
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1986 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1987 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]1988 return 0;
1989
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1990 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1991 .addOperand(Dest).addOperand(Src), -1);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1992 break;
1993 }
1994 case X86::DEC16r:
1995 case X86::DEC64_16r:
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]1996 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]1997 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]1998 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]1999 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2000 .addOperand(Dest).addOperand(Src), -1);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2001 break;
2002 case X86::ADD64rr:
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]2003 case X86::ADD64rr_DB:
2004 case X86::ADD32rr:
2005 case X86::ADD32rr_DB: {
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2006 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]2007 unsigned Opc;
Craig Topper44d23822012-02-22 05:59:10 +0000[diff] [blame]2008 const TargetRegisterClass *RC;
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]2009 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
2010 Opc = X86::LEA64r;
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]2011 RC = &X86::GR64_NOSPRegClass;
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]2012 } else {
2013 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]2014 RC = &X86::GR32_NOSPRegClass;
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]2015 }
2016
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2017
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2018 unsigned Src2 = MI->getOperand(2).getReg();
2019 bool isKill2 = MI->getOperand(2).isKill();
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]2020
2021 // LEA can't handle RSP.
2022 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]2023 !MF.getRegInfo().constrainRegClass(Src2, RC))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +0000[diff] [blame]2024 return 0;
2025
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]2026 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]2027 .addOperand(Dest),
2028 Src.getReg(), Src.isKill(), Src2, isKill2);
Nadav Rotemd93ea882012-07-16 10:52:25 +0000[diff] [blame]2029
2030 // Preserve undefness of the operands.
2031 bool isUndef = MI->getOperand(1).isUndef();
2032 bool isUndef2 = MI->getOperand(2).isUndef();
2033 NewMI->getOperand(1).setIsUndef(isUndef);
2034 NewMI->getOperand(3).setIsUndef(isUndef2);
2035
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2036 if (LV && isKill2)
2037 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2038 break;
2039 }
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]2040 case X86::ADD16rr:
2041 case X86::ADD16rr_DB: {
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2042 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]2043 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2044 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2045 unsigned Src2 = MI->getOperand(2).getReg();
2046 bool isKill2 = MI->getOperand(2).isKill();
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]2047 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]2048 .addOperand(Dest),
2049 Src.getReg(), Src.isKill(), Src2, isKill2);
2050
2051 // Preserve undefness of the operands.
2052 bool isUndef = MI->getOperand(1).isUndef();
2053 bool isUndef2 = MI->getOperand(2).isUndef();
2054 NewMI->getOperand(1).setIsUndef(isUndef);
2055 NewMI->getOperand(3).setIsUndef(isUndef2);
2056
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2057 if (LV && isKill2)
2058 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2059 break;
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2060 }
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2061 case X86::ADD64ri32:
2062 case X86::ADD64ri8:
Chris Lattner15df55d2010-10-08 03:57:25 +0000[diff] [blame]2063 case X86::ADD64ri32_DB:
2064 case X86::ADD64ri8_DB:
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2065 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]2066 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2067 .addOperand(Dest).addOperand(Src),
2068 MI->getOperand(2).getImm());
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2069 break;
2070 case X86::ADD32ri:
Chris Lattner15df55d2010-10-08 03:57:25 +0000[diff] [blame]2071 case X86::ADD32ri8:
2072 case X86::ADD32ri_DB:
2073 case X86::ADD32ri8_DB: {
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2074 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2075 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]2076 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2077 .addOperand(Dest).addOperand(Src),
2078 MI->getOperand(2).getImm());
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2079 break;
2080 }
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2081 case X86::ADD16ri:
2082 case X86::ADD16ri8:
Chris Lattner15df55d2010-10-08 03:57:25 +0000[diff] [blame]2083 case X86::ADD16ri_DB:
2084 case X86::ADD16ri8_DB:
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2085 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +0000[diff] [blame]2086 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2087 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]2088 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2089 .addOperand(Dest).addOperand(Src),
2090 MI->getOperand(2).getImm());
Evan Cheng656e5142009-12-11 06:01:48 +0000[diff] [blame]2091 break;
Evan Cheng559dc462007-10-05 20:34:26 +0000[diff] [blame]2092 }
2093 }
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]2094 }
2095
Evan Cheng15246732008-02-07 08:29:53 +0000[diff] [blame]2096 if (!NewMI) return 0;
2097
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2098 if (LV) { // Update live variables
Jakob Stoklund Olesen91f3a6c2012-08-23 22:36:31 +0000[diff] [blame]2099 if (Src.isKill())
2100 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2101 if (Dest.isDead())
2102 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2103 }
2104
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2105 MFI->insert(MBBI, NewMI); // Insert the new inst
Evan Cheng6ce7dc22006-11-15 20:58:11 +0000[diff] [blame]2106 return NewMI;
Chris Lattnerbcea4d62005-01-02 02:37:07 +0000[diff] [blame]2107}
2108
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2109/// commuteInstruction - We have a few instructions that must be hacked on to
2110/// commute them.
2111///
Evan Cheng58dcb0e2008-06-16 07:33:11 +0000[diff] [blame]2112MachineInstr *
2113X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2114 switch (MI->getOpcode()) {
Chris Lattner0df53d22005-01-19 07:31:24 +0000[diff] [blame]2115 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2116 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2117 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
Dan Gohmane47f1f92007-09-14 23:17:45 +0000[diff] [blame]2118 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2119 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2120 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
Chris Lattner0df53d22005-01-19 07:31:24 +0000[diff] [blame]2121 unsigned Opc;
2122 unsigned Size;
2123 switch (MI->getOpcode()) {
Torok Edwinc23197a2009-07-14 16:55:14 +0000[diff] [blame]2124 default: llvm_unreachable("Unreachable!");
Chris Lattner0df53d22005-01-19 07:31:24 +0000[diff] [blame]2125 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2126 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2127 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2128 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
Dan Gohmane47f1f92007-09-14 23:17:45 +0000[diff] [blame]2129 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2130 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
Chris Lattner0df53d22005-01-19 07:31:24 +0000[diff] [blame]2131 }
Chris Lattner9a1ceae2007-12-30 20:49:49 +0000[diff] [blame]2132 unsigned Amt = MI->getOperand(3).getImm();
Dan Gohman74feef22008-10-17 01:23:35 +0000[diff] [blame]2133 if (NewMI) {
2134 MachineFunction &MF = *MI->getParent()->getParent();
2135 MI = MF.CloneMachineInstr(MI);
2136 NewMI = false;
Evan Chenga4d16a12008-02-13 02:46:49 +0000[diff] [blame]2137 }
Dan Gohman74feef22008-10-17 01:23:35 +0000[diff] [blame]2138 MI->setDesc(get(Opc));
2139 MI->getOperand(3).setImm(Size-Amt);
2140 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2141 }
Craig Toppercba48d82012-08-21 07:32:16 +0000[diff] [blame]2142 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2143 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2144 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2145 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2146 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2147 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2148 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2149 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2150 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2151 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2152 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2153 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2154 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2155 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2156 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2157 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2158 unsigned Opc;
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2159 switch (MI->getOpcode()) {
Craig Toppercba48d82012-08-21 07:32:16 +0000[diff] [blame]2160 default: llvm_unreachable("Unreachable!");
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2161 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2162 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2163 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2164 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2165 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2166 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2167 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2168 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2169 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2170 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2171 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2172 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
Chris Lattner25cbf502010-10-05 23:00:14 +0000[diff] [blame]2173 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2174 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2175 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2176 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2177 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2178 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2179 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2180 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2181 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2182 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2183 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2184 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2185 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2186 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2187 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2188 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2189 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2190 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2191 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2192 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
Mon P Wang0bd07fc2009-04-18 05:16:01 +0000[diff] [blame]2193 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2194 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2195 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2196 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2197 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2198 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
Mon P Wang0bd07fc2009-04-18 05:16:01 +0000[diff] [blame]2199 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2200 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2201 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2202 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
Dan Gohman305fceb2009-01-07 00:35:10 +0000[diff] [blame]2203 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2204 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
Mon P Wang0bd07fc2009-04-18 05:16:01 +0000[diff] [blame]2205 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
Dan Gohman305fceb2009-01-07 00:35:10 +0000[diff] [blame]2206 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2207 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2208 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2209 }
Dan Gohman74feef22008-10-17 01:23:35 +0000[diff] [blame]2210 if (NewMI) {
2211 MachineFunction &MF = *MI->getParent()->getParent();
2212 MI = MF.CloneMachineInstr(MI);
2213 NewMI = false;
2214 }
Chris Lattner5080f4d2008-01-11 18:10:50 +0000[diff] [blame]2215 MI->setDesc(get(Opc));
Evan Cheng7ad42d92007-10-05 23:13:21 +0000[diff] [blame]2216 // Fallthrough intended.
2217 }
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2218 default:
Evan Cheng58dcb0e2008-06-16 07:33:11 +0000[diff] [blame]2219 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner41e431b2005-01-19 07:11:01 +0000[diff] [blame]2220 }
2221}
2222
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2223static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2224 switch (BrOpc) {
2225 default: return X86::COND_INVALID;
Chris Lattnerbd13fb62010-02-11 19:25:55 +0000[diff] [blame]2226 case X86::JE_4: return X86::COND_E;
2227 case X86::JNE_4: return X86::COND_NE;
2228 case X86::JL_4: return X86::COND_L;
2229 case X86::JLE_4: return X86::COND_LE;
2230 case X86::JG_4: return X86::COND_G;
2231 case X86::JGE_4: return X86::COND_GE;
2232 case X86::JB_4: return X86::COND_B;
2233 case X86::JBE_4: return X86::COND_BE;
2234 case X86::JA_4: return X86::COND_A;
2235 case X86::JAE_4: return X86::COND_AE;
2236 case X86::JS_4: return X86::COND_S;
2237 case X86::JNS_4: return X86::COND_NS;
2238 case X86::JP_4: return X86::COND_P;
2239 case X86::JNP_4: return X86::COND_NP;
2240 case X86::JO_4: return X86::COND_O;
2241 case X86::JNO_4: return X86::COND_NO;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2242 }
2243}
2244
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2245/// getCondFromSETOpc - return condition code of a SET opcode.
2246static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2247 switch (Opc) {
2248 default: return X86::COND_INVALID;
2249 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2250 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2251 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2252 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2253 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2254 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2255 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2256 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2257 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2258 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2259 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2260 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2261 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2262 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2263 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2264 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2265 }
2266}
2267
2268/// getCondFromCmovOpc - return condition code of a CMov opcode.
2269static X86::CondCode getCondFromCMovOpc(unsigned Opc) {
2270 switch (Opc) {
2271 default: return X86::COND_INVALID;
2272 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2273 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2274 return X86::COND_A;
2275 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2276 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2277 return X86::COND_AE;
2278 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2279 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2280 return X86::COND_B;
2281 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2282 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2283 return X86::COND_BE;
2284 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2285 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2286 return X86::COND_E;
2287 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2288 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2289 return X86::COND_G;
2290 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2291 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2292 return X86::COND_GE;
2293 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2294 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2295 return X86::COND_L;
2296 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2297 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2298 return X86::COND_LE;
2299 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2300 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2301 return X86::COND_NE;
2302 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2303 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2304 return X86::COND_NO;
2305 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2306 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2307 return X86::COND_NP;
2308 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2309 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2310 return X86::COND_NS;
2311 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2312 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2313 return X86::COND_O;
2314 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2315 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2316 return X86::COND_P;
2317 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2318 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2319 return X86::COND_S;
2320 }
2321}
2322
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2323unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2324 switch (CC) {
Torok Edwinc23197a2009-07-14 16:55:14 +0000[diff] [blame]2325 default: llvm_unreachable("Illegal condition code!");
Chris Lattnerbd13fb62010-02-11 19:25:55 +0000[diff] [blame]2326 case X86::COND_E: return X86::JE_4;
2327 case X86::COND_NE: return X86::JNE_4;
2328 case X86::COND_L: return X86::JL_4;
2329 case X86::COND_LE: return X86::JLE_4;
2330 case X86::COND_G: return X86::JG_4;
2331 case X86::COND_GE: return X86::JGE_4;
2332 case X86::COND_B: return X86::JB_4;
2333 case X86::COND_BE: return X86::JBE_4;
2334 case X86::COND_A: return X86::JA_4;
2335 case X86::COND_AE: return X86::JAE_4;
2336 case X86::COND_S: return X86::JS_4;
2337 case X86::COND_NS: return X86::JNS_4;
2338 case X86::COND_P: return X86::JP_4;
2339 case X86::COND_NP: return X86::JNP_4;
2340 case X86::COND_O: return X86::JO_4;
2341 case X86::COND_NO: return X86::JNO_4;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2342 }
2343}
2344
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]2345/// GetOppositeBranchCondition - Return the inverse of the specified condition,
2346/// e.g. turning COND_E to COND_NE.
2347X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2348 switch (CC) {
Torok Edwinc23197a2009-07-14 16:55:14 +0000[diff] [blame]2349 default: llvm_unreachable("Illegal condition code!");
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]2350 case X86::COND_E: return X86::COND_NE;
2351 case X86::COND_NE: return X86::COND_E;
2352 case X86::COND_L: return X86::COND_GE;
2353 case X86::COND_LE: return X86::COND_G;
2354 case X86::COND_G: return X86::COND_LE;
2355 case X86::COND_GE: return X86::COND_L;
2356 case X86::COND_B: return X86::COND_AE;
2357 case X86::COND_BE: return X86::COND_A;
2358 case X86::COND_A: return X86::COND_BE;
2359 case X86::COND_AE: return X86::COND_B;
2360 case X86::COND_S: return X86::COND_NS;
2361 case X86::COND_NS: return X86::COND_S;
2362 case X86::COND_P: return X86::COND_NP;
2363 case X86::COND_NP: return X86::COND_P;
2364 case X86::COND_O: return X86::COND_NO;
2365 case X86::COND_NO: return X86::COND_O;
2366 }
2367}
2368
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2369/// getSwappedCondition - assume the flags are set by MI(a,b), return
2370/// the condition code if we modify the instructions such that flags are
2371/// set by MI(b,a).
Benjamin Kramer23d36222012-07-13 13:25:15 +0000[diff] [blame]2372static X86::CondCode getSwappedCondition(X86::CondCode CC) {
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2373 switch (CC) {
2374 default: return X86::COND_INVALID;
2375 case X86::COND_E: return X86::COND_E;
2376 case X86::COND_NE: return X86::COND_NE;
2377 case X86::COND_L: return X86::COND_G;
2378 case X86::COND_LE: return X86::COND_GE;
2379 case X86::COND_G: return X86::COND_L;
2380 case X86::COND_GE: return X86::COND_LE;
2381 case X86::COND_B: return X86::COND_A;
2382 case X86::COND_BE: return X86::COND_AE;
2383 case X86::COND_A: return X86::COND_B;
2384 case X86::COND_AE: return X86::COND_BE;
2385 }
2386}
2387
2388/// getSETFromCond - Return a set opcode for the given condition and
2389/// whether it has memory operand.
2390static unsigned getSETFromCond(X86::CondCode CC,
2391 bool HasMemoryOperand) {
Craig Toppera1823672012-08-21 08:23:21 +0000[diff] [blame]2392 static const uint16_t Opc[16][2] = {
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2393 { X86::SETAr, X86::SETAm },
2394 { X86::SETAEr, X86::SETAEm },
2395 { X86::SETBr, X86::SETBm },
2396 { X86::SETBEr, X86::SETBEm },
2397 { X86::SETEr, X86::SETEm },
2398 { X86::SETGr, X86::SETGm },
2399 { X86::SETGEr, X86::SETGEm },
2400 { X86::SETLr, X86::SETLm },
2401 { X86::SETLEr, X86::SETLEm },
2402 { X86::SETNEr, X86::SETNEm },
2403 { X86::SETNOr, X86::SETNOm },
2404 { X86::SETNPr, X86::SETNPm },
2405 { X86::SETNSr, X86::SETNSm },
2406 { X86::SETOr, X86::SETOm },
2407 { X86::SETPr, X86::SETPm },
2408 { X86::SETSr, X86::SETSm }
2409 };
2410
2411 assert(CC < 16 && "Can only handle standard cond codes");
2412 return Opc[CC][HasMemoryOperand ? 1 : 0];
2413}
2414
2415/// getCMovFromCond - Return a cmov opcode for the given condition,
2416/// register size in bytes, and operand type.
2417static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2418 bool HasMemoryOperand) {
Craig Toppera1823672012-08-21 08:23:21 +0000[diff] [blame]2419 static const uint16_t Opc[32][3] = {
Jakob Stoklund Olesen59bde4d2012-07-04 00:09:58 +0000[diff] [blame]2420 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2421 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2422 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2423 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2424 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2425 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2426 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2427 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2428 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2429 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2430 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2431 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2432 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2433 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2434 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2435 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2436 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2437 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2438 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2439 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2440 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2441 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2442 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2443 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2444 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2445 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2446 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2447 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2448 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2449 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2450 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2451 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
Jakob Stoklund Olesen59bde4d2012-07-04 00:09:58 +0000[diff] [blame]2452 };
2453
2454 assert(CC < 16 && "Can only handle standard cond codes");
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2455 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
Jakob Stoklund Olesen59bde4d2012-07-04 00:09:58 +0000[diff] [blame]2456 switch(RegBytes) {
2457 default: llvm_unreachable("Illegal register size!");
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2458 case 2: return Opc[Idx][0];
2459 case 4: return Opc[Idx][1];
2460 case 8: return Opc[Idx][2];
Jakob Stoklund Olesen59bde4d2012-07-04 00:09:58 +0000[diff] [blame]2461 }
2462}
2463
Dale Johannesen318093b2007-06-14 22:03:45 +0000[diff] [blame]2464bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
Evan Cheng5a96b3d2011-12-07 07:15:52 +0000[diff] [blame]2465 if (!MI->isTerminator()) return false;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2466
Chris Lattner69244302008-01-07 01:56:04 +0000[diff] [blame]2467 // Conditional branch is a special case.
Evan Cheng5a96b3d2011-12-07 07:15:52 +0000[diff] [blame]2468 if (MI->isBranch() && !MI->isBarrier())
Chris Lattner69244302008-01-07 01:56:04 +0000[diff] [blame]2469 return true;
Evan Cheng5a96b3d2011-12-07 07:15:52 +0000[diff] [blame]2470 if (!MI->isPredicable())
Chris Lattner69244302008-01-07 01:56:04 +0000[diff] [blame]2471 return true;
2472 return !isPredicated(MI);
Dale Johannesen318093b2007-06-14 22:03:45 +0000[diff] [blame]2473}
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]2474
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2475bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2476 MachineBasicBlock *&TBB,
2477 MachineBasicBlock *&FBB,
Evan Chengdc54d312009-02-09 07:14:22 +0000[diff] [blame]2478 SmallVectorImpl<MachineOperand> &Cond,
2479 bool AllowModify) const {
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2480 // Start from the bottom of the block and work up, examining the
2481 // terminator instructions.
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2482 MachineBasicBlock::iterator I = MBB.end();
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2483 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2484 while (I != MBB.begin()) {
2485 --I;
Dale Johannesen93d6a7e2010-04-02 01:38:09 +0000[diff] [blame]2486 if (I->isDebugValue())
2487 continue;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2488
2489 // Working from the bottom, when we see a non-terminator instruction, we're
2490 // done.
Jakob Stoklund Olesen468a2a42010-07-16 17:41:44 +0000[diff] [blame]2491 if (!isUnpredicatedTerminator(I))
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2492 break;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2493
2494 // A terminator that isn't a branch can't easily be handled by this
2495 // analysis.
Evan Cheng5a96b3d2011-12-07 07:15:52 +0000[diff] [blame]2496 if (!I->isBranch())
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2497 return true;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2498
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2499 // Handle unconditional branches.
Chris Lattnerbd13fb62010-02-11 19:25:55 +0000[diff] [blame]2500 if (I->getOpcode() == X86::JMP_4) {
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2501 UnCondBrIter = I;
2502
Evan Chengdc54d312009-02-09 07:14:22 +0000[diff] [blame]2503 if (!AllowModify) {
2504 TBB = I->getOperand(0).getMBB();
Evan Cheng45e00102009-05-08 06:34:09 +0000[diff] [blame]2505 continue;
Evan Chengdc54d312009-02-09 07:14:22 +0000[diff] [blame]2506 }
2507
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2508 // If the block has any instructions after a JMP, delete them.
Chris Lattner7896c9f2009-12-03 00:50:42 +0000[diff] [blame]2509 while (llvm::next(I) != MBB.end())
2510 llvm::next(I)->eraseFromParent();
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2511
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2512 Cond.clear();
2513 FBB = 0;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2514
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2515 // Delete the JMP if it's equivalent to a fall-through.
2516 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2517 TBB = 0;
2518 I->eraseFromParent();
2519 I = MBB.end();
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2520 UnCondBrIter = MBB.end();
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2521 continue;
2522 }
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2523
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2524 // TBB is used to indicate the unconditional destination.
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2525 TBB = I->getOperand(0).getMBB();
2526 continue;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2527 }
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2528
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2529 // Handle conditional branches.
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2530 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2531 if (BranchCode == X86::COND_INVALID)
2532 return true; // Can't handle indirect branch.
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2533
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2534 // Working from the bottom, handle the first conditional branch.
2535 if (Cond.empty()) {
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2536 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2537 if (AllowModify && UnCondBrIter != MBB.end() &&
2538 MBB.isLayoutSuccessor(TargetBB)) {
2539 // If we can modify the code and it ends in something like:
2540 //
2541 // jCC L1
2542 // jmp L2
2543 // L1:
2544 // ...
2545 // L2:
2546 //
2547 // Then we can change this to:
2548 //
2549 // jnCC L2
2550 // L1:
2551 // ...
2552 // L2:
2553 //
2554 // Which is a bit more efficient.
2555 // We conditionally jump to the fall-through block.
2556 BranchCode = GetOppositeBranchCondition(BranchCode);
2557 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2558 MachineBasicBlock::iterator OldInst = I;
2559
2560 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2561 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2562 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2563 .addMBB(TargetBB);
Evan Chengfc5a03e2010-04-13 18:50:27 +0000[diff] [blame]2564
2565 OldInst->eraseFromParent();
2566 UnCondBrIter->eraseFromParent();
2567
2568 // Restart the analysis.
2569 UnCondBrIter = MBB.end();
2570 I = MBB.end();
2571 continue;
2572 }
2573
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2574 FBB = TBB;
2575 TBB = I->getOperand(0).getMBB();
2576 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2577 continue;
2578 }
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2579
2580 // Handle subsequent conditional branches. Only handle the case where all
2581 // conditional branches branch to the same destination and their condition
2582 // opcodes fit one of the special multi-branch idioms.
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2583 assert(Cond.size() == 1);
2584 assert(TBB);
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2585
2586 // Only handle the case where all conditional branches branch to the same
2587 // destination.
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2588 if (TBB != I->getOperand(0).getMBB())
2589 return true;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2590
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2591 // If the conditions are the same, we can leave them alone.
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2592 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2593 if (OldBranchCode == BranchCode)
2594 continue;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2595
2596 // If they differ, see if they fit one of the known patterns. Theoretically,
2597 // we could handle more patterns here, but we shouldn't expect to see them
2598 // if instruction selection has done a reasonable job.
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2599 if ((OldBranchCode == X86::COND_NP &&
2600 BranchCode == X86::COND_E) ||
2601 (OldBranchCode == X86::COND_E &&
2602 BranchCode == X86::COND_NP))
2603 BranchCode = X86::COND_NP_OR_E;
2604 else if ((OldBranchCode == X86::COND_P &&
2605 BranchCode == X86::COND_NE) ||
2606 (OldBranchCode == X86::COND_NE &&
2607 BranchCode == X86::COND_P))
2608 BranchCode = X86::COND_NE_OR_P;
2609 else
2610 return true;
Bill Wendling85de1e52009-12-14 06:51:19 +0000[diff] [blame]2611
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2612 // Update the MachineOperand.
2613 Cond[0].setImm(BranchCode);
Chris Lattner6ce64432006-10-30 22:27:23 +0000[diff] [blame]2614 }
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2615
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2616 return false;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2617}
2618
Evan Cheng6ae36262007-05-18 00:18:17 +0000[diff] [blame]2619unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2620 MachineBasicBlock::iterator I = MBB.end();
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2621 unsigned Count = 0;
2622
2623 while (I != MBB.begin()) {
2624 --I;
Dale Johannesen93d6a7e2010-04-02 01:38:09 +0000[diff] [blame]2625 if (I->isDebugValue())
2626 continue;
Chris Lattnerbd13fb62010-02-11 19:25:55 +0000[diff] [blame]2627 if (I->getOpcode() != X86::JMP_4 &&
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2628 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2629 break;
2630 // Remove the branch.
2631 I->eraseFromParent();
2632 I = MBB.end();
2633 ++Count;
2634 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]2635
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2636 return Count;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2637}
2638
Evan Cheng6ae36262007-05-18 00:18:17 +0000[diff] [blame]2639unsigned
2640X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2641 MachineBasicBlock *FBB,
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2642 const SmallVectorImpl<MachineOperand> &Cond,
2643 DebugLoc DL) const {
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2644 // Shouldn't be a fall through.
2645 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
Chris Lattner34a84ac2006-10-21 05:34:23 +0000[diff] [blame]2646 assert((Cond.size() == 1 || Cond.size() == 0) &&
2647 "X86 branch conditions have one component!");
2648
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2649 if (Cond.empty()) {
2650 // Unconditional branch?
2651 assert(!FBB && "Unconditional branch with multiple successors!");
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2652 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
Evan Cheng6ae36262007-05-18 00:18:17 +0000[diff] [blame]2653 return 1;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2654 }
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2655
2656 // Conditional branch.
2657 unsigned Count = 0;
2658 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2659 switch (CC) {
2660 case X86::COND_NP_OR_E:
2661 // Synthesize NP_OR_E with two branches.
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2662 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2663 ++Count;
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2664 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2665 ++Count;
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2666 break;
2667 case X86::COND_NE_OR_P:
2668 // Synthesize NE_OR_P with two branches.
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2669 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2670 ++Count;
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2671 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2672 ++Count;
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2673 break;
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2674 default: {
2675 unsigned Opc = GetCondBranchFromCond(CC);
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2676 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2677 ++Count;
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2678 }
Bill Wendling18ce64e2010-03-05 00:33:59 +0000[diff] [blame]2679 }
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2680 if (FBB) {
2681 // Two-way Conditional branch. Insert the second branch.
Stuart Hastings3bf91252010-06-17 22:43:56 +0000[diff] [blame]2682 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]2683 ++Count;
2684 }
2685 return Count;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]2686}
2687
Jakob Stoklund Olesen59bde4d2012-07-04 00:09:58 +0000[diff] [blame]2688bool X86InstrInfo::
2689canInsertSelect(const MachineBasicBlock &MBB,
2690 const SmallVectorImpl<MachineOperand> &Cond,
2691 unsigned TrueReg, unsigned FalseReg,
2692 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2693 // Not all subtargets have cmov instructions.
2694 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2695 return false;
2696 if (Cond.size() != 1)
2697 return false;
2698 // We cannot do the composite conditions, at least not in SSA form.
2699 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2700 return false;
2701
2702 // Check register classes.
2703 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2704 const TargetRegisterClass *RC =
2705 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2706 if (!RC)
2707 return false;
2708
2709 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2710 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2711 X86::GR32RegClass.hasSubClassEq(RC) ||
2712 X86::GR64RegClass.hasSubClassEq(RC)) {
2713 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2714 // Bridge. Probably Ivy Bridge as well.
2715 CondCycles = 2;
2716 TrueCycles = 2;
2717 FalseCycles = 2;
2718 return true;
2719 }
2720
2721 // Can't do vectors.
2722 return false;
2723}
2724
2725void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2726 MachineBasicBlock::iterator I, DebugLoc DL,
2727 unsigned DstReg,
2728 const SmallVectorImpl<MachineOperand> &Cond,
2729 unsigned TrueReg, unsigned FalseReg) const {
2730 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2731 assert(Cond.size() == 1 && "Invalid Cond array");
2732 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]2733 MRI.getRegClass(DstReg)->getSize(),
2734 false/*HasMemoryOperand*/);
Jakob Stoklund Olesen59bde4d2012-07-04 00:09:58 +0000[diff] [blame]2735 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2736}
2737
Dan Gohman6d9305c2009-04-15 00:04:23 +0000[diff] [blame]2738/// isHReg - Test if the given register is a physical h register.
2739static bool isHReg(unsigned Reg) {
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2740 return X86::GR8_ABCD_HRegClass.contains(Reg);
Dan Gohman6d9305c2009-04-15 00:04:23 +0000[diff] [blame]2741}
2742
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2743// Try and copy between VR128/VR64 and GR64 registers.
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2744static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2745 bool HasAVX) {
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2746 // SrcReg(VR128) -> DestReg(GR64)
2747 // SrcReg(VR64) -> DestReg(GR64)
2748 // SrcReg(GR64) -> DestReg(VR128)
2749 // SrcReg(GR64) -> DestReg(VR64)
2750
2751 if (X86::GR64RegClass.contains(DestReg)) {
Craig Topper4dea9062012-08-21 08:29:51 +0000[diff] [blame]2752 if (X86::VR128RegClass.contains(SrcReg))
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2753 // Copy from a VR128 register to a GR64 register.
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2754 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
Craig Topper4dea9062012-08-21 08:29:51 +0000[diff] [blame]2755 if (X86::VR64RegClass.contains(SrcReg))
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2756 // Copy from a VR64 register to a GR64 register.
2757 return X86::MOVSDto64rr;
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2758 } else if (X86::GR64RegClass.contains(SrcReg)) {
2759 // Copy from a GR64 register to a VR128 register.
2760 if (X86::VR128RegClass.contains(DestReg))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2761 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2762 // Copy from a GR64 register to a VR64 register.
Craig Topper4dea9062012-08-21 08:29:51 +0000[diff] [blame]2763 if (X86::VR64RegClass.contains(DestReg))
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2764 return X86::MOV64toSDrr;
2765 }
2766
Jakob Stoklund Olesen4bd89872011-09-22 22:45:24 +0000[diff] [blame]2767 // SrcReg(FR32) -> DestReg(GR32)
2768 // SrcReg(GR32) -> DestReg(FR32)
2769
2770 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
Craig Topper4dea9062012-08-21 08:29:51 +0000[diff] [blame]2771 // Copy from a FR32 register to a GR32 register.
2772 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
Jakob Stoklund Olesen4bd89872011-09-22 22:45:24 +0000[diff] [blame]2773
2774 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
Craig Topper4dea9062012-08-21 08:29:51 +0000[diff] [blame]2775 // Copy from a GR32 register to a FR32 register.
2776 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
Jakob Stoklund Olesen4bd89872011-09-22 22:45:24 +0000[diff] [blame]2777
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2778 return 0;
2779}
2780
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2781void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2782 MachineBasicBlock::iterator MI, DebugLoc DL,
2783 unsigned DestReg, unsigned SrcReg,
2784 bool KillSrc) const {
2785 // First deal with the normal symmetric copies.
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2786 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Craig Topper4dea9062012-08-21 08:29:51 +0000[diff] [blame]2787 unsigned Opc;
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2788 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2789 Opc = X86::MOV64rr;
2790 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2791 Opc = X86::MOV32rr;
2792 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2793 Opc = X86::MOV16rr;
2794 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2795 // Copying to or from a physical H register on x86-64 requires a NOREX
2796 // move. Otherwise use a normal move.
2797 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
Jakob Stoklund Olesenb66f1842011-10-07 20:15:54 +0000[diff] [blame]2798 TM.getSubtarget<X86Subtarget>().is64Bit()) {
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2799 Opc = X86::MOV8rr_NOREX;
Jakob Stoklund Olesenb66f1842011-10-07 20:15:54 +0000[diff] [blame]2800 // Both operands must be encodable without an REX prefix.
2801 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2802 "8-bit H register can not be copied outside GR8_NOREX");
2803 } else
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2804 Opc = X86::MOV8rr;
2805 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2806 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]2807 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2808 Opc = X86::VMOVAPSYrr;
Jakob Stoklund Olesen61c8ecc2010-07-08 22:30:35 +0000[diff] [blame]2809 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2810 Opc = X86::MMX_MOVQ64rr;
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +0000[diff] [blame]2811 else
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2812 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2813
2814 if (Opc) {
2815 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2816 .addReg(SrcReg, getKillRegState(KillSrc));
2817 return;
2818 }
2819
2820 // Moving EFLAGS to / from another register requires a push and a pop.
2821 if (SrcReg == X86::EFLAGS) {
2822 if (X86::GR64RegClass.contains(DestReg)) {
2823 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2824 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2825 return;
Craig Topper4dea9062012-08-21 08:29:51 +0000[diff] [blame]2826 }
2827 if (X86::GR32RegClass.contains(DestReg)) {
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2828 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2829 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2830 return;
2831 }
2832 }
2833 if (DestReg == X86::EFLAGS) {
2834 if (X86::GR64RegClass.contains(SrcReg)) {
2835 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2836 .addReg(SrcReg, getKillRegState(KillSrc));
2837 BuildMI(MBB, MI, DL, get(X86::POPF64));
2838 return;
Craig Topper4dea9062012-08-21 08:29:51 +0000[diff] [blame]2839 }
2840 if (X86::GR32RegClass.contains(SrcReg)) {
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +0000[diff] [blame]2841 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2842 .addReg(SrcReg, getKillRegState(KillSrc));
2843 BuildMI(MBB, MI, DL, get(X86::POPF32));
2844 return;
2845 }
2846 }
2847
2848 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2849 << " to " << RI.getName(DestReg) << '\n');
2850 llvm_unreachable("Cannot emit physreg copy instruction");
2851}
2852
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2853static unsigned getLoadStoreRegOpcode(unsigned Reg,
2854 const TargetRegisterClass *RC,
2855 bool isStackAligned,
2856 const TargetMachine &TM,
2857 bool load) {
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2858 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2859 switch (RC->getSize()) {
Rafael Espindola5a717a32010-07-12 03:43:04 +0000[diff] [blame]2860 default:
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2861 llvm_unreachable("Unknown spill size");
2862 case 1:
2863 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2864 if (TM.getSubtarget<X86Subtarget>().is64Bit())
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2865 // Copying to or from a physical H register on x86-64 requires a NOREX
2866 // move. Otherwise use a normal move.
2867 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2868 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2869 return load ? X86::MOV8rm : X86::MOV8mr;
2870 case 2:
2871 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2872 return load ? X86::MOV16rm : X86::MOV16mr;
2873 case 4:
2874 if (X86::GR32RegClass.hasSubClassEq(RC))
2875 return load ? X86::MOV32rm : X86::MOV32mr;
2876 if (X86::FR32RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2877 return load ?
2878 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2879 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2880 if (X86::RFP32RegClass.hasSubClassEq(RC))
2881 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2882 llvm_unreachable("Unknown 4-byte regclass");
2883 case 8:
2884 if (X86::GR64RegClass.hasSubClassEq(RC))
2885 return load ? X86::MOV64rm : X86::MOV64mr;
2886 if (X86::FR64RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2887 return load ?
2888 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
2889 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2890 if (X86::VR64RegClass.hasSubClassEq(RC))
2891 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2892 if (X86::RFP64RegClass.hasSubClassEq(RC))
2893 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2894 llvm_unreachable("Unknown 8-byte regclass");
2895 case 10:
2896 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2897 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
Bruno Cardoso Lopes5affa512011-08-31 03:04:09 +0000[diff] [blame]2898 case 16: {
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +0000[diff] [blame]2899 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2900 // If stack is realigned we can use aligned stores.
2901 if (isStackAligned)
Bruno Cardoso Lopes5affa512011-08-31 03:04:09 +0000[diff] [blame]2902 return load ?
2903 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
2904 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2905 else
Bruno Cardoso Lopes5affa512011-08-31 03:04:09 +0000[diff] [blame]2906 return load ?
2907 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
2908 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
2909 }
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]2910 case 32:
2911 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2912 // If stack is realigned we can use aligned stores.
2913 if (isStackAligned)
2914 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2915 else
2916 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2917 }
2918}
2919
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2920static unsigned getStoreRegOpcode(unsigned SrcReg,
2921 const TargetRegisterClass *RC,
2922 bool isStackAligned,
2923 TargetMachine &TM) {
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2924 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2925}
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2926
Rafael Espindola21d238f2010-06-12 20:13:29 +0000[diff] [blame]2927
2928static unsigned getLoadRegOpcode(unsigned DestReg,
2929 const TargetRegisterClass *RC,
2930 bool isStackAligned,
2931 const TargetMachine &TM) {
2932 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2933}
2934
2935void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
2936 MachineBasicBlock::iterator MI,
2937 unsigned SrcReg, bool isKill, int FrameIdx,
Evan Cheng746ad692010-05-06 19:06:44 +0000[diff] [blame]2938 const TargetRegisterClass *RC,
2939 const TargetRegisterInfo *TRI) const {
Anton Korobeynikov88bbf692008-07-19 06:30:51 +0000[diff] [blame]2940 const MachineFunction &MF = *MBB.getParent();
Jakob Stoklund Olesen516cd452010-07-27 04:16:58 +0000[diff] [blame]2941 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
2942 "Stack slot too small for store");
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2943 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2944 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Cheng2fa82bc2011-06-23 01:53:43 +0000[diff] [blame]2945 RI.canRealignStack(MF);
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2946 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Dale Johannesen6ec25f52010-01-26 00:03:12 +0000[diff] [blame]2947 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]2948 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]2949 .addReg(SrcReg, getKillRegState(isKill));
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2950}
2951
2952void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
2953 bool isKill,
2954 SmallVectorImpl<MachineOperand> &Addr,
2955 const TargetRegisterClass *RC,
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]2956 MachineInstr::mmo_iterator MMOBegin,
2957 MachineInstr::mmo_iterator MMOEnd,
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2958 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2959 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2960 bool isAligned = MMOBegin != MMOEnd &&
2961 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2962 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Chris Lattnerc7f3ace2010-04-02 20:16:16 +0000[diff] [blame]2963 DebugLoc DL;
Dale Johannesen21b55412009-02-12 23:08:38 +0000[diff] [blame]2964 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2965 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]2966 MIB.addOperand(Addr[i]);
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]2967 MIB.addReg(SrcReg, getKillRegState(isKill));
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]2968 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2969 NewMIs.push_back(MIB);
2970}
2971
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2972
2973void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
Anton Korobeynikov88bbf692008-07-19 06:30:51 +0000[diff] [blame]2974 MachineBasicBlock::iterator MI,
2975 unsigned DestReg, int FrameIdx,
Evan Cheng746ad692010-05-06 19:06:44 +0000[diff] [blame]2976 const TargetRegisterClass *RC,
2977 const TargetRegisterInfo *TRI) const {
Anton Korobeynikov88bbf692008-07-19 06:30:51 +0000[diff] [blame]2978 const MachineFunction &MF = *MBB.getParent();
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2979 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2980 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Cheng2fa82bc2011-06-23 01:53:43 +0000[diff] [blame]2981 RI.canRealignStack(MF);
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2982 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Dale Johannesen6ec25f52010-01-26 00:03:12 +0000[diff] [blame]2983 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]2984 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2985}
2986
2987void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]2988 SmallVectorImpl<MachineOperand> &Addr,
2989 const TargetRegisterClass *RC,
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]2990 MachineInstr::mmo_iterator MMOBegin,
2991 MachineInstr::mmo_iterator MMOEnd,
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2992 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]2993 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2994 bool isAligned = MMOBegin != MMOEnd &&
2995 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman4af325d2009-04-27 16:41:36 +0000[diff] [blame]2996 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Chris Lattnerc7f3ace2010-04-02 20:16:16 +0000[diff] [blame]2997 DebugLoc DL;
Dale Johannesen21b55412009-02-12 23:08:38 +0000[diff] [blame]2998 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]2999 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3000 MIB.addOperand(Addr[i]);
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]3001 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersonf6372aa2008-01-01 21:11:32 +0000[diff] [blame]3002 NewMIs.push_back(MIB);
3003}
3004
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3005bool X86InstrInfo::
3006analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3007 int &CmpMask, int &CmpValue) const {
3008 switch (MI->getOpcode()) {
3009 default: break;
3010 case X86::CMP64ri32:
3011 case X86::CMP64ri8:
3012 case X86::CMP32ri:
3013 case X86::CMP32ri8:
3014 case X86::CMP16ri:
3015 case X86::CMP16ri8:
3016 case X86::CMP8ri:
3017 SrcReg = MI->getOperand(0).getReg();
3018 SrcReg2 = 0;
3019 CmpMask = ~0;
3020 CmpValue = MI->getOperand(1).getImm();
3021 return true;
Manman Ren39ad5682012-08-08 00:51:41 +0000[diff] [blame]3022 // A SUB can be used to perform comparison.
3023 case X86::SUB64rm:
3024 case X86::SUB32rm:
3025 case X86::SUB16rm:
3026 case X86::SUB8rm:
3027 SrcReg = MI->getOperand(1).getReg();
3028 SrcReg2 = 0;
3029 CmpMask = ~0;
3030 CmpValue = 0;
3031 return true;
3032 case X86::SUB64rr:
3033 case X86::SUB32rr:
3034 case X86::SUB16rr:
3035 case X86::SUB8rr:
3036 SrcReg = MI->getOperand(1).getReg();
3037 SrcReg2 = MI->getOperand(2).getReg();
3038 CmpMask = ~0;
3039 CmpValue = 0;
3040 return true;
3041 case X86::SUB64ri32:
3042 case X86::SUB64ri8:
3043 case X86::SUB32ri:
3044 case X86::SUB32ri8:
3045 case X86::SUB16ri:
3046 case X86::SUB16ri8:
3047 case X86::SUB8ri:
3048 SrcReg = MI->getOperand(1).getReg();
3049 SrcReg2 = 0;
3050 CmpMask = ~0;
3051 CmpValue = MI->getOperand(2).getImm();
3052 return true;
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3053 case X86::CMP64rr:
3054 case X86::CMP32rr:
3055 case X86::CMP16rr:
3056 case X86::CMP8rr:
3057 SrcReg = MI->getOperand(0).getReg();
3058 SrcReg2 = MI->getOperand(1).getReg();
3059 CmpMask = ~0;
3060 CmpValue = 0;
3061 return true;
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3062 case X86::TEST8rr:
3063 case X86::TEST16rr:
3064 case X86::TEST32rr:
3065 case X86::TEST64rr:
3066 SrcReg = MI->getOperand(0).getReg();
3067 if (MI->getOperand(1).getReg() != SrcReg) return false;
3068 // Compare against zero.
3069 SrcReg2 = 0;
3070 CmpMask = ~0;
3071 CmpValue = 0;
3072 return true;
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3073 }
3074 return false;
3075}
3076
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3077/// isRedundantFlagInstr - check whether the first instruction, whose only
3078/// purpose is to update flags, can be made redundant.
3079/// CMPrr can be made redundant by SUBrr if the operands are the same.
3080/// This function can be extended later on.
3081/// SrcReg, SrcRegs: register operands for FlagI.
3082/// ImmValue: immediate for FlagI if it takes an immediate.
3083inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3084 unsigned SrcReg2, int ImmValue,
3085 MachineInstr *OI) {
3086 if (((FlagI->getOpcode() == X86::CMP64rr &&
3087 OI->getOpcode() == X86::SUB64rr) ||
3088 (FlagI->getOpcode() == X86::CMP32rr &&
3089 OI->getOpcode() == X86::SUB32rr)||
3090 (FlagI->getOpcode() == X86::CMP16rr &&
3091 OI->getOpcode() == X86::SUB16rr)||
3092 (FlagI->getOpcode() == X86::CMP8rr &&
3093 OI->getOpcode() == X86::SUB8rr)) &&
3094 ((OI->getOperand(1).getReg() == SrcReg &&
3095 OI->getOperand(2).getReg() == SrcReg2) ||
3096 (OI->getOperand(1).getReg() == SrcReg2 &&
3097 OI->getOperand(2).getReg() == SrcReg)))
3098 return true;
3099
3100 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3101 OI->getOpcode() == X86::SUB64ri32) ||
3102 (FlagI->getOpcode() == X86::CMP64ri8 &&
3103 OI->getOpcode() == X86::SUB64ri8) ||
3104 (FlagI->getOpcode() == X86::CMP32ri &&
3105 OI->getOpcode() == X86::SUB32ri) ||
3106 (FlagI->getOpcode() == X86::CMP32ri8 &&
3107 OI->getOpcode() == X86::SUB32ri8) ||
3108 (FlagI->getOpcode() == X86::CMP16ri &&
3109 OI->getOpcode() == X86::SUB16ri) ||
3110 (FlagI->getOpcode() == X86::CMP16ri8 &&
3111 OI->getOpcode() == X86::SUB16ri8) ||
3112 (FlagI->getOpcode() == X86::CMP8ri &&
3113 OI->getOpcode() == X86::SUB8ri)) &&
3114 OI->getOperand(1).getReg() == SrcReg &&
3115 OI->getOperand(2).getImm() == ImmValue)
3116 return true;
3117 return false;
3118}
3119
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3120/// isDefConvertible - check whether the definition can be converted
3121/// to remove a comparison against zero.
3122inline static bool isDefConvertible(MachineInstr *MI) {
3123 switch (MI->getOpcode()) {
3124 default: return false;
3125 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3126 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3127 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3128 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3129 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3130 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3131 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3132 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3133 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3134 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3135 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3136 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3137 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3138 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3139 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3140 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3141 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3142 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3143 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3144 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3145 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3146 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3147 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3148 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3149 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3150 return true;
3151 }
3152}
3153
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3154/// optimizeCompareInstr - Check if there exists an earlier instruction that
3155/// operates on the same source operands and sets flags in the same way as
3156/// Compare; remove Compare if possible.
3157bool X86InstrInfo::
3158optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3159 int CmpMask, int CmpValue,
3160 const MachineRegisterInfo *MRI) const {
Manman Ren39ad5682012-08-08 00:51:41 +0000[diff] [blame]3161 // Check whether we can replace SUB with CMP.
3162 unsigned NewOpcode = 0;
3163 switch (CmpInstr->getOpcode()) {
3164 default: break;
3165 case X86::SUB64ri32:
3166 case X86::SUB64ri8:
3167 case X86::SUB32ri:
3168 case X86::SUB32ri8:
3169 case X86::SUB16ri:
3170 case X86::SUB16ri8:
3171 case X86::SUB8ri:
3172 case X86::SUB64rm:
3173 case X86::SUB32rm:
3174 case X86::SUB16rm:
3175 case X86::SUB8rm:
3176 case X86::SUB64rr:
3177 case X86::SUB32rr:
3178 case X86::SUB16rr:
3179 case X86::SUB8rr: {
3180 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3181 return false;
3182 // There is no use of the destination register, we can replace SUB with CMP.
3183 switch (CmpInstr->getOpcode()) {
Craig Topper195f1b82012-08-21 08:16:16 +0000[diff] [blame]3184 default: llvm_unreachable("Unreachable!");
Manman Ren39ad5682012-08-08 00:51:41 +0000[diff] [blame]3185 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3186 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3187 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3188 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3189 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3190 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3191 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3192 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3193 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3194 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3195 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3196 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3197 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3198 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3199 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3200 }
3201 CmpInstr->setDesc(get(NewOpcode));
3202 CmpInstr->RemoveOperand(0);
3203 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3204 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3205 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3206 return false;
3207 }
3208 }
3209
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3210 // Get the unique definition of SrcReg.
3211 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3212 if (!MI) return false;
3213
3214 // CmpInstr is the first instruction of the BB.
3215 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3216
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3217 // If we are comparing against zero, check whether we can use MI to update
3218 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3219 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3220 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3221 !isDefConvertible(MI)))
3222 return false;
3223
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3224 // We are searching for an earlier instruction that can make CmpInstr
3225 // redundant and that instruction will be saved in Sub.
3226 MachineInstr *Sub = NULL;
3227 const TargetRegisterInfo *TRI = &getRegisterInfo();
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]3228
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3229 // We iterate backward, starting from the instruction before CmpInstr and
3230 // stop when reaching the definition of a source register or done with the BB.
3231 // RI points to the instruction before CmpInstr.
3232 // If the definition is in this basic block, RE points to the definition;
3233 // otherwise, RE is the rend of the basic block.
3234 MachineBasicBlock::reverse_iterator
3235 RI = MachineBasicBlock::reverse_iterator(I),
3236 RE = CmpInstr->getParent() == MI->getParent() ?
3237 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3238 CmpInstr->getParent()->rend();
Manman Ren84ae7e92012-07-11 19:35:12 +0000[diff] [blame]3239 MachineInstr *Movr0Inst = 0;
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3240 for (; RI != RE; ++RI) {
3241 MachineInstr *Instr = &*RI;
3242 // Check whether CmpInstr can be made redundant by the current instruction.
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3243 if (!IsCmpZero &&
3244 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3245 Sub = Instr;
3246 break;
3247 }
3248
3249 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
Manman Ren84ae7e92012-07-11 19:35:12 +0000[diff] [blame]3250 Instr->readsRegister(X86::EFLAGS, TRI)) {
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3251 // This instruction modifies or uses EFLAGS.
Manman Ren84ae7e92012-07-11 19:35:12 +0000[diff] [blame]3252
3253 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3254 // They are safe to move up, if the definition to EFLAGS is dead and
3255 // earlier instructions do not read or write EFLAGS.
3256 if (!Movr0Inst && (Instr->getOpcode() == X86::MOV8r0 ||
3257 Instr->getOpcode() == X86::MOV16r0 ||
3258 Instr->getOpcode() == X86::MOV32r0 ||
3259 Instr->getOpcode() == X86::MOV64r0) &&
3260 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3261 Movr0Inst = Instr;
3262 continue;
3263 }
3264
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3265 // We can't remove CmpInstr.
3266 return false;
Manman Ren84ae7e92012-07-11 19:35:12 +0000[diff] [blame]3267 }
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3268 }
3269
3270 // Return false if no candidates exist.
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3271 if (!IsCmpZero && !Sub)
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3272 return false;
3273
Manman Ren2d4215f2012-07-07 03:34:46 +0000[diff] [blame]3274 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3275 Sub->getOperand(2).getReg() == SrcReg);
3276
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3277 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
Manman Ren2d4215f2012-07-07 03:34:46 +0000[diff] [blame]3278 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3279 // If we are done with the basic block, we need to check whether EFLAGS is
3280 // live-out.
3281 bool IsSafe = false;
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3282 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3283 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3284 for (++I; I != E; ++I) {
3285 const MachineInstr &Instr = *I;
Manman Ren43d9ab12012-07-28 03:15:46 +0000[diff] [blame]3286 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3287 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3288 // We should check the usage if this instruction uses and updates EFLAGS.
3289 if (!UseEFLAGS && ModifyEFLAGS) {
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3290 // It is safe to remove CmpInstr if EFLAGS is updated again.
Manman Ren2d4215f2012-07-07 03:34:46 +0000[diff] [blame]3291 IsSafe = true;
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3292 break;
Manman Ren2d4215f2012-07-07 03:34:46 +0000[diff] [blame]3293 }
Manman Ren43d9ab12012-07-28 03:15:46 +0000[diff] [blame]3294 if (!UseEFLAGS && !ModifyEFLAGS)
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3295 continue;
3296
3297 // EFLAGS is used by this instruction.
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3298 X86::CondCode OldCC;
3299 bool OpcIsSET = false;
3300 if (IsCmpZero || IsSwapped) {
3301 // We decode the condition code from opcode.
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]3302 if (Instr.isBranch())
3303 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3304 else {
3305 OldCC = getCondFromSETOpc(Instr.getOpcode());
3306 if (OldCC != X86::COND_INVALID)
3307 OpcIsSET = true;
3308 else
3309 OldCC = getCondFromCMovOpc(Instr.getOpcode());
3310 }
3311 if (OldCC == X86::COND_INVALID) return false;
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3312 }
3313 if (IsCmpZero) {
3314 switch (OldCC) {
3315 default: break;
3316 case X86::COND_A: case X86::COND_AE:
3317 case X86::COND_B: case X86::COND_BE:
3318 case X86::COND_G: case X86::COND_GE:
3319 case X86::COND_L: case X86::COND_LE:
3320 case X86::COND_O: case X86::COND_NO:
3321 // CF and OF are used, we can't perform this optimization.
3322 return false;
3323 }
3324 } else if (IsSwapped) {
3325 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3326 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3327 // We swap the condition code and synthesize the new opcode.
Manman Ren62093642012-07-09 18:57:12 +0000[diff] [blame]3328 X86::CondCode NewCC = getSwappedCondition(OldCC);
3329 if (NewCC == X86::COND_INVALID) return false;
3330
3331 // Synthesize the new opcode.
3332 bool HasMemoryOperand = Instr.hasOneMemOperand();
3333 unsigned NewOpc;
3334 if (Instr.isBranch())
3335 NewOpc = GetCondBranchFromCond(NewCC);
3336 else if(OpcIsSET)
3337 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3338 else {
3339 unsigned DstReg = Instr.getOperand(0).getReg();
3340 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3341 HasMemoryOperand);
3342 }
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3343
3344 // Push the MachineInstr to OpsToUpdate.
3345 // If it is safe to remove CmpInstr, the condition code of these
3346 // instructions will be modified.
3347 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3348 }
Manman Ren43d9ab12012-07-28 03:15:46 +0000[diff] [blame]3349 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3350 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
Manman Ren2d4215f2012-07-07 03:34:46 +0000[diff] [blame]3351 IsSafe = true;
3352 break;
3353 }
3354 }
3355
3356 // If EFLAGS is not killed nor re-defined, we should check whether it is
3357 // live-out. If it is live-out, do not optimize.
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3358 if ((IsCmpZero || IsSwapped) && !IsSafe) {
Manman Ren2d4215f2012-07-07 03:34:46 +0000[diff] [blame]3359 MachineBasicBlock *MBB = CmpInstr->getParent();
3360 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3361 SE = MBB->succ_end(); SI != SE; ++SI)
3362 if ((*SI)->isLiveIn(X86::EFLAGS))
3363 return false;
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3364 }
3365
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3366 // The instruction to be updated is either Sub or MI.
3367 Sub = IsCmpZero ? MI : Sub;
Manman Ren84ae7e92012-07-11 19:35:12 +0000[diff] [blame]3368 // Move Movr0Inst to the place right before Sub.
3369 if (Movr0Inst) {
3370 Sub->getParent()->remove(Movr0Inst);
3371 Sub->getParent()->insert(MachineBasicBlock::iterator(Sub), Movr0Inst);
3372 }
3373
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3374 // Make sure Sub instruction defines EFLAGS.
Manman Ren62a89f52012-07-18 21:40:01 +0000[diff] [blame]3375 assert(Sub->getNumOperands() >= 2 &&
3376 Sub->getOperand(Sub->getNumOperands()-1).isReg() &&
3377 Sub->getOperand(Sub->getNumOperands()-1).getReg() == X86::EFLAGS &&
3378 "EFLAGS should be the last operand of SUB, ADD, OR, XOR, AND");
3379 Sub->getOperand(Sub->getNumOperands()-1).setIsDef(true);
Manman Ren2af66dc2012-07-06 17:36:20 +0000[diff] [blame]3380 CmpInstr->eraseFromParent();
3381
3382 // Modify the condition code of instructions in OpsToUpdate.
3383 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3384 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3385 return true;
3386}
3387
Manman Rend7d003c2012-08-02 00:56:42 +0000[diff] [blame]3388/// optimizeLoadInstr - Try to remove the load by folding it to a register
3389/// operand at the use. We fold the load instructions if load defines a virtual
3390/// register, the virtual register is used once in the same BB, and the
3391/// instructions in-between do not load or store, and have no side effects.
3392MachineInstr* X86InstrInfo::
3393optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3394 unsigned &FoldAsLoadDefReg,
3395 MachineInstr *&DefMI) const {
3396 if (FoldAsLoadDefReg == 0)
3397 return 0;
3398 // To be conservative, if there exists another load, clear the load candidate.
3399 if (MI->mayLoad()) {
3400 FoldAsLoadDefReg = 0;
3401 return 0;
3402 }
3403
3404 // Check whether we can move DefMI here.
3405 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3406 assert(DefMI);
3407 bool SawStore = false;
3408 if (!DefMI->isSafeToMove(this, 0, SawStore))
3409 return 0;
3410
3411 // We try to commute MI if possible.
3412 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3413 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3414 // Collect information about virtual register operands of MI.
3415 unsigned SrcOperandId = 0;
3416 bool FoundSrcOperand = false;
3417 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3418 MachineOperand &MO = MI->getOperand(i);
3419 if (!MO.isReg())
3420 continue;
3421 unsigned Reg = MO.getReg();
3422 if (Reg != FoldAsLoadDefReg)
3423 continue;
3424 // Do not fold if we have a subreg use or a def or multiple uses.
3425 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3426 return 0;
3427
3428 SrcOperandId = i;
3429 FoundSrcOperand = true;
3430 }
3431 if (!FoundSrcOperand) return 0;
3432
3433 // Check whether we can fold the def into SrcOperandId.
3434 SmallVector<unsigned, 8> Ops;
3435 Ops.push_back(SrcOperandId);
3436 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3437 if (FoldMI) {
3438 FoldAsLoadDefReg = 0;
3439 return FoldMI;
3440 }
3441
3442 if (Idx == 1) {
3443 // MI was changed but it didn't help, commute it back!
3444 commuteInstruction(MI, false);
3445 return 0;
3446 }
3447
3448 // Check whether we can commute MI and enable folding.
3449 if (MI->isCommutable()) {
3450 MachineInstr *NewMI = commuteInstruction(MI, false);
3451 // Unable to commute.
3452 if (!NewMI) return 0;
3453 if (NewMI != MI) {
3454 // New instruction. It doesn't need to be kept.
3455 NewMI->eraseFromParent();
3456 return 0;
3457 }
3458 }
3459 }
3460 return 0;
3461}
3462
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3463/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3464/// instruction with two undef reads of the register being defined. This is
3465/// used for mapping:
3466/// %xmm4 = V_SET0
3467/// to:
3468/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3469///
3470static bool Expand2AddrUndef(MachineInstr *MI, const MCInstrDesc &Desc) {
3471 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3472 unsigned Reg = MI->getOperand(0).getReg();
3473 MI->setDesc(Desc);
3474
3475 // MachineInstr::addOperand() will insert explicit operands before any
3476 // implicit operands.
3477 MachineInstrBuilder(MI).addReg(Reg, RegState::Undef)
3478 .addReg(Reg, RegState::Undef);
3479 // But we don't trust that.
3480 assert(MI->getOperand(1).getReg() == Reg &&
3481 MI->getOperand(2).getReg() == Reg && "Misplaced operand");
3482 return true;
3483}
3484
3485bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3486 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3487 switch (MI->getOpcode()) {
3488 case X86::V_SET0:
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +0000[diff] [blame]3489 case X86::FsFLD0SS:
3490 case X86::FsFLD0SD:
Jakob Stoklund Olesen3e5d5c52011-11-07 19:15:58 +0000[diff] [blame]3491 return Expand2AddrUndef(MI, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
Craig Topper13897fb2012-08-28 07:05:28 +0000[diff] [blame]3492 case X86::AVX_SET0:
3493 assert(HasAVX && "AVX not supported");
3494 return Expand2AddrUndef(MI, get(X86::VXORPSYrr));
Craig Topperd9021942012-08-28 07:30:47 +0000[diff] [blame]3495 case X86::V_SETALLONES:
3496 return Expand2AddrUndef(MI, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
3497 case X86::AVX2_SETALLONES:
3498 return Expand2AddrUndef(MI, get(X86::VPCMPEQDYrr));
Jakob Stoklund Olesened744822011-10-08 18:28:28 +0000[diff] [blame]3499 case X86::TEST8ri_NOREX:
3500 MI->setDesc(get(X86::TEST8ri));
3501 return true;
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3502 }
3503 return false;
3504}
3505
Evan Cheng962021b2010-04-26 07:38:55 +0000[diff] [blame]3506MachineInstr*
3507X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
Evan Cheng8601a3d2010-04-29 01:13:30 +0000[diff] [blame]3508 int FrameIx, uint64_t Offset,
Evan Cheng962021b2010-04-26 07:38:55 +0000[diff] [blame]3509 const MDNode *MDPtr,
3510 DebugLoc DL) const {
Evan Cheng962021b2010-04-26 07:38:55 +0000[diff] [blame]3511 X86AddressMode AM;
3512 AM.BaseType = X86AddressMode::FrameIndexBase;
3513 AM.Base.FrameIndex = FrameIx;
3514 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
3515 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
3516 return &*MIB;
3517}
3518
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]3519static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
Dan Gohmand68a0762009-01-05 17:59:02 +0000[diff] [blame]3520 const SmallVectorImpl<MachineOperand> &MOs,
Bill Wendling9bc96a52009-02-03 00:55:04 +0000[diff] [blame]3521 MachineInstr *MI,
3522 const TargetInstrInfo &TII) {
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3523 // Create the base instruction with the memory operand as the first part.
Bill Wendling9bc96a52009-02-03 00:55:04 +0000[diff] [blame]3524 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3525 MI->getDebugLoc(), true);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3526 MachineInstrBuilder MIB(NewMI);
3527 unsigned NumAddrOps = MOs.size();
3528 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3529 MIB.addOperand(MOs[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3530 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]3531 addOffset(MIB, 0);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3532
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3533 // Loop over the rest of the ri operands, converting them over.
Chris Lattner749c6f62008-01-07 07:27:27 +0000[diff] [blame]3534 unsigned NumOps = MI->getDesc().getNumOperands()-2;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3535 for (unsigned i = 0; i != NumOps; ++i) {
3536 MachineOperand &MO = MI->getOperand(i+2);
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3537 MIB.addOperand(MO);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3538 }
3539 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3540 MachineOperand &MO = MI->getOperand(i);
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3541 MIB.addOperand(MO);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3542 }
3543 return MIB;
3544}
3545
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]3546static MachineInstr *FuseInst(MachineFunction &MF,
3547 unsigned Opcode, unsigned OpNo,
Dan Gohmand68a0762009-01-05 17:59:02 +0000[diff] [blame]3548 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3549 MachineInstr *MI, const TargetInstrInfo &TII) {
Bill Wendling9bc96a52009-02-03 00:55:04 +0000[diff] [blame]3550 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3551 MI->getDebugLoc(), true);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3552 MachineInstrBuilder MIB(NewMI);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3553
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3554 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3555 MachineOperand &MO = MI->getOperand(i);
3556 if (i == OpNo) {
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]3557 assert(MO.isReg() && "Expected to fold into reg operand!");
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3558 unsigned NumAddrOps = MOs.size();
3559 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3560 MIB.addOperand(MOs[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3561 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]3562 addOffset(MIB, 0);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3563 } else {
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3564 MIB.addOperand(MO);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3565 }
3566 }
3567 return MIB;
3568}
3569
3570static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
Dan Gohmand68a0762009-01-05 17:59:02 +0000[diff] [blame]3571 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3572 MachineInstr *MI) {
Dan Gohman8e5f2c62008-07-07 23:14:23 +0000[diff] [blame]3573 MachineFunction &MF = *MI->getParent()->getParent();
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]3574 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3575
3576 unsigned NumAddrOps = MOs.size();
3577 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]3578 MIB.addOperand(MOs[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3579 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]3580 addOffset(MIB, 0);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3581 return MIB.addImm(0);
3582}
3583
3584MachineInstr*
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3585X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3586 MachineInstr *MI, unsigned i,
Evan Chengf9b36f02009-07-15 06:10:07 +0000[diff] [blame]3587 const SmallVectorImpl<MachineOperand> &MOs,
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3588 unsigned Size, unsigned Align) const {
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3589 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3590 bool isTwoAddrFold = false;
Chris Lattner749c6f62008-01-07 07:27:27 +0000[diff] [blame]3591 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3592 bool isTwoAddr = NumOps > 1 &&
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3593 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3594
Jakob Stoklund Olesen60045c22011-04-30 23:00:05 +0000[diff] [blame]3595 // FIXME: AsmPrinter doesn't know how to handle
3596 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3597 if (MI->getOpcode() == X86::ADD32ri &&
3598 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3599 return NULL;
3600
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3601 MachineInstr *NewMI = NULL;
3602 // Folding a memory location into the two-address part of a two-address
3603 // instruction is different than folding it other places. It requires
3604 // replacing the *two* registers with the memory location.
3605 if (isTwoAddr && NumOps >= 2 && i < 2 &&
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]3606 MI->getOperand(0).isReg() &&
3607 MI->getOperand(1).isReg() &&
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3608 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3609 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3610 isTwoAddrFold = true;
3611 } else if (i == 0) { // If operand 0
Craig Topper8a5bc5a2012-08-23 04:57:36 +0000[diff] [blame]3612 unsigned Opc = 0;
3613 switch (MI->getOpcode()) {
3614 default: break;
3615 case X86::MOV64r0: Opc = X86::MOV64mi32; break;
3616 case X86::MOV32r0: Opc = X86::MOV32mi; break;
3617 case X86::MOV16r0: Opc = X86::MOV16mi; break;
3618 case X86::MOV8r0: Opc = X86::MOV8mi; break;
3619 }
3620 if (Opc)
3621 NewMI = MakeM0Inst(*this, Opc, MOs, MI);
Evan Cheng9f1c8312008-07-03 09:09:37 +0000[diff] [blame]3622 if (NewMI)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3623 return NewMI;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3624
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3625 OpcodeTablePtr = &RegOp2MemOpTable0;
3626 } else if (i == 1) {
3627 OpcodeTablePtr = &RegOp2MemOpTable1;
3628 } else if (i == 2) {
3629 OpcodeTablePtr = &RegOp2MemOpTable2;
Elena Demikhovsky1503aba2012-08-01 12:06:00 +0000[diff] [blame]3630 } else if (i == 3) {
3631 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3632 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3633
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3634 // If table selected...
3635 if (OpcodeTablePtr) {
3636 // Find the Opcode to fuse
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3637 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3638 OpcodeTablePtr->find(MI->getOpcode());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3639 if (I != OpcodeTablePtr->end()) {
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3640 unsigned Opcode = I->second.first;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]3641 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
Evan Chengf9b36f02009-07-15 06:10:07 +0000[diff] [blame]3642 if (Align < MinAlign)
3643 return NULL;
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3644 bool NarrowToMOV32rm = false;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3645 if (Size) {
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]3646 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3647 if (Size < RCSize) {
3648 // Check if it's safe to fold the load. If the size of the object is
3649 // narrower than the load width, then it's not.
3650 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3651 return NULL;
3652 // If this is a 64-bit load, but the spill slot is 32, then we can do
3653 // a 32-bit load which is implicitly zero-extended. This likely is due
3654 // to liveintervalanalysis remat'ing a load from stack slot.
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3655 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3656 return NULL;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3657 Opcode = X86::MOV32rm;
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3658 NarrowToMOV32rm = true;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3659 }
3660 }
3661
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3662 if (isTwoAddrFold)
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3663 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3664 else
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3665 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3666
3667 if (NarrowToMOV32rm) {
3668 // If this is the special case where we use a MOV32rm to load a 32-bit
3669 // value and zero-extend the top bits. Change the destination register
3670 // to a 32-bit one.
3671 unsigned DstReg = NewMI->getOperand(0).getReg();
3672 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3673 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
Jakob Stoklund Olesen3458e9e2010-05-24 14:48:17 +0000[diff] [blame]3674 X86::sub_32bit));
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3675 else
Jakob Stoklund Olesen3458e9e2010-05-24 14:48:17 +0000[diff] [blame]3676 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
Evan Cheng879caea2009-09-11 01:01:31 +0000[diff] [blame]3677 }
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3678 return NewMI;
3679 }
3680 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3681
3682 // No fusion
Jakob Stoklund Olesen9c50e8b2010-07-09 20:43:09 +0000[diff] [blame]3683 if (PrintFailedFusing && !MI->isCopy())
David Greene5b901322010-01-05 01:29:29 +0000[diff] [blame]3684 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3685 return NULL;
3686}
3687
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3688/// hasPartialRegUpdate - Return true for all instructions that only update
3689/// the first 32 or 64-bits of the destination register and leave the rest
3690/// unmodified. This can be used to avoid folding loads if the instructions
3691/// only update part of the destination register, and the non-updated part is
3692/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3693/// instructions breaks the partial register dependency and it can improve
3694/// performance. e.g.:
3695///
3696/// movss (%rdi), %xmm0
3697/// cvtss2sd %xmm0, %xmm0
3698///
3699/// Instead of
3700/// cvtss2sd (%rdi), %xmm0
3701///
Bruno Cardoso Lopes6b5b79c2011-09-15 23:04:24 +0000[diff] [blame]3702/// FIXME: This should be turned into a TSFlags.
3703///
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3704static bool hasPartialRegUpdate(unsigned Opcode) {
3705 switch (Opcode) {
Jakob Stoklund Olesenc2ecf3e2011-11-15 01:15:30 +0000[diff] [blame]3706 case X86::CVTSI2SSrr:
3707 case X86::CVTSI2SS64rr:
3708 case X86::CVTSI2SDrr:
3709 case X86::CVTSI2SD64rr:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3710 case X86::CVTSD2SSrr:
3711 case X86::Int_CVTSD2SSrr:
3712 case X86::CVTSS2SDrr:
3713 case X86::Int_CVTSS2SDrr:
3714 case X86::RCPSSr:
3715 case X86::RCPSSr_Int:
3716 case X86::ROUNDSDr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +0000[diff] [blame]3717 case X86::ROUNDSDr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3718 case X86::ROUNDSSr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +0000[diff] [blame]3719 case X86::ROUNDSSr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3720 case X86::RSQRTSSr:
3721 case X86::RSQRTSSr_Int:
3722 case X86::SQRTSSr:
3723 case X86::SQRTSSr_Int:
3724 // AVX encoded versions
3725 case X86::VCVTSD2SSrr:
3726 case X86::Int_VCVTSD2SSrr:
3727 case X86::VCVTSS2SDrr:
3728 case X86::Int_VCVTSS2SDrr:
3729 case X86::VRCPSSr:
3730 case X86::VROUNDSDr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +0000[diff] [blame]3731 case X86::VROUNDSDr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3732 case X86::VROUNDSSr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +0000[diff] [blame]3733 case X86::VROUNDSSr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3734 case X86::VRSQRTSSr:
3735 case X86::VSQRTSSr:
3736 return true;
3737 }
3738
3739 return false;
3740}
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3741
Jakob Stoklund Olesenc2ecf3e2011-11-15 01:15:30 +0000[diff] [blame]3742/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3743/// instructions we would like before a partial register update.
3744unsigned X86InstrInfo::
3745getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3746 const TargetRegisterInfo *TRI) const {
3747 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3748 return 0;
3749
3750 // If MI is marked as reading Reg, the partial register update is wanted.
3751 const MachineOperand &MO = MI->getOperand(0);
3752 unsigned Reg = MO.getReg();
3753 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3754 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3755 return 0;
3756 } else {
3757 if (MI->readsRegister(Reg, TRI))
3758 return 0;
3759 }
3760
3761 // If any of the preceding 16 instructions are reading Reg, insert a
3762 // dependency breaking instruction. The magic number is based on a few
3763 // Nehalem experiments.
3764 return 16;
3765}
3766
3767void X86InstrInfo::
3768breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3769 const TargetRegisterInfo *TRI) const {
3770 unsigned Reg = MI->getOperand(OpNum).getReg();
3771 if (X86::VR128RegClass.contains(Reg)) {
3772 // These instructions are all floating point domain, so xorps is the best
3773 // choice.
3774 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3775 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3776 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3777 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3778 } else if (X86::VR256RegClass.contains(Reg)) {
3779 // Use vxorps to clear the full ymm register.
3780 // It wants to read and write the xmm sub-register.
3781 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3782 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3783 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3784 .addReg(Reg, RegState::ImplicitDefine);
3785 } else
3786 return;
3787 MI->addRegisterKilled(Reg, TRI, true);
3788}
3789
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3790MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3791 MachineInstr *MI,
Evan Chengf9b36f02009-07-15 06:10:07 +0000[diff] [blame]3792 const SmallVectorImpl<unsigned> &Ops,
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3793 int FrameIndex) const {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3794 // Check switch flag
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3795 if (NoFusing) return NULL;
3796
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3797 // Unless optimizing for size, don't fold to avoid partial
3798 // register update stalls
3799 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3800 hasPartialRegUpdate(MI->getOpcode()))
3801 return 0;
Evan Cheng400073d2009-12-18 07:40:29 +0000[diff] [blame]3802
Evan Cheng5fd79d02008-02-08 21:20:40 +0000[diff] [blame]3803 const MachineFrameInfo *MFI = MF.getFrameInfo();
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3804 unsigned Size = MFI->getObjectSize(FrameIndex);
Evan Cheng5fd79d02008-02-08 21:20:40 +0000[diff] [blame]3805 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3806 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3807 unsigned NewOpc = 0;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3808 unsigned RCSize = 0;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3809 switch (MI->getOpcode()) {
3810 default: return NULL;
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3811 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
Dan Gohmane5efbaf2010-05-18 21:42:03 +0000[diff] [blame]3812 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
3813 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
3814 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3815 }
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3816 // Check if it's safe to fold the load. If the size of the object is
3817 // narrower than the load width, then it's not.
3818 if (Size < RCSize)
3819 return NULL;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3820 // Change to CMPXXri r, 0 first.
Chris Lattner5080f4d2008-01-11 18:10:50 +0000[diff] [blame]3821 MI->setDesc(get(NewOpc));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3822 MI->getOperand(1).ChangeToImmediate(0);
3823 } else if (Ops.size() != 1)
3824 return NULL;
3825
3826 SmallVector<MachineOperand,4> MOs;
3827 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3828 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3829}
3830
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3831MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3832 MachineInstr *MI,
Evan Chengf9b36f02009-07-15 06:10:07 +0000[diff] [blame]3833 const SmallVectorImpl<unsigned> &Ops,
Dan Gohmanc54baa22008-12-03 18:43:12 +0000[diff] [blame]3834 MachineInstr *LoadMI) const {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3835 // Check switch flag
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3836 if (NoFusing) return NULL;
3837
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +0000[diff] [blame]3838 // Unless optimizing for size, don't fold to avoid partial
3839 // register update stalls
3840 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3841 hasPartialRegUpdate(MI->getOpcode()))
3842 return 0;
Evan Cheng400073d2009-12-18 07:40:29 +0000[diff] [blame]3843
Dan Gohmancddc11e2008-07-12 00:10:52 +0000[diff] [blame]3844 // Determine the alignment of the load.
Evan Cheng5fd79d02008-02-08 21:20:40 +0000[diff] [blame]3845 unsigned Alignment = 0;
Dan Gohmancddc11e2008-07-12 00:10:52 +0000[diff] [blame]3846 if (LoadMI->hasOneMemOperand())
Dan Gohmanc76909a2009-09-25 20:36:54 +0000[diff] [blame]3847 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3848 else
3849 switch (LoadMI->getOpcode()) {
Craig Topper745a86b2011-11-19 22:34:59 +0000[diff] [blame]3850 case X86::AVX2_SETALLONES:
Craig Topper13897fb2012-08-28 07:05:28 +0000[diff] [blame]3851 case X86::AVX_SET0:
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]3852 Alignment = 32;
3853 break;
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3854 case X86::V_SET0:
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3855 case X86::V_SETALLONES:
3856 Alignment = 16;
3857 break;
3858 case X86::FsFLD0SD:
3859 Alignment = 8;
3860 break;
3861 case X86::FsFLD0SS:
3862 Alignment = 4;
3863 break;
3864 default:
Eli Friedmanbe5cbaa2011-06-10 01:13:01 +0000[diff] [blame]3865 return 0;
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3866 }
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3867 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3868 unsigned NewOpc = 0;
3869 switch (MI->getOpcode()) {
3870 default: return NULL;
3871 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]3872 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
3873 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
3874 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3875 }
3876 // Change to CMPXXri r, 0 first.
Chris Lattner5080f4d2008-01-11 18:10:50 +0000[diff] [blame]3877 MI->setDesc(get(NewOpc));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3878 MI->getOperand(1).ChangeToImmediate(0);
3879 } else if (Ops.size() != 1)
3880 return NULL;
3881
Jakob Stoklund Olesend29583b2010-08-11 23:08:22 +0000[diff] [blame]3882 // Make sure the subregisters match.
3883 // Otherwise we risk changing the size of the load.
3884 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
3885 return NULL;
3886
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]3887 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3888 switch (LoadMI->getOpcode()) {
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3889 case X86::V_SET0:
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3890 case X86::V_SETALLONES:
Craig Topper745a86b2011-11-19 22:34:59 +0000[diff] [blame]3891 case X86::AVX2_SETALLONES:
Craig Topper13897fb2012-08-28 07:05:28 +0000[diff] [blame]3892 case X86::AVX_SET0:
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3893 case X86::FsFLD0SD:
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +0000[diff] [blame]3894 case X86::FsFLD0SS: {
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +0000[diff] [blame]3895 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3896 // Create a constant-pool entry and operands to load from it.
3897
Dan Gohman81d0c362010-03-09 03:01:40 +0000[diff] [blame]3898 // Medium and large mode can't fold loads this way.
3899 if (TM.getCodeModel() != CodeModel::Small &&
3900 TM.getCodeModel() != CodeModel::Kernel)
3901 return NULL;
3902
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3903 // x86-32 PIC requires a PIC base register for constant pools.
3904 unsigned PICBase = 0;
Jakob Stoklund Olesen93e55de2009-07-16 21:24:13 +0000[diff] [blame]3905 if (TM.getRelocationModel() == Reloc::PIC_) {
Evan Cheng2b48ab92009-07-16 18:44:05 +0000[diff] [blame]3906 if (TM.getSubtarget<X86Subtarget>().is64Bit())
3907 PICBase = X86::RIP;
Jakob Stoklund Olesen93e55de2009-07-16 21:24:13 +0000[diff] [blame]3908 else
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]3909 // FIXME: PICBase = getGlobalBaseReg(&MF);
Evan Cheng2b48ab92009-07-16 18:44:05 +0000[diff] [blame]3910 // This doesn't work for several reasons.
3911 // 1. GlobalBaseReg may have been spilled.
3912 // 2. It may not be live at MI.
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3913 return NULL;
Jakob Stoklund Olesen93e55de2009-07-16 21:24:13 +0000[diff] [blame]3914 }
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3915
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3916 // Create a constant-pool entry.
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3917 MachineConstantPool &MCP = *MF.getConstantPool();
Chris Lattnerdb125cf2011-07-18 04:54:35 +0000[diff] [blame]3918 Type *Ty;
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]3919 unsigned Opc = LoadMI->getOpcode();
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +0000[diff] [blame]3920 if (Opc == X86::FsFLD0SS)
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3921 Ty = Type::getFloatTy(MF.getFunction()->getContext());
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +0000[diff] [blame]3922 else if (Opc == X86::FsFLD0SD)
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3923 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
Craig Topper13897fb2012-08-28 07:05:28 +0000[diff] [blame]3924 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
Craig Topperb9c7f652012-01-13 06:12:41 +0000[diff] [blame]3925 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3926 else
3927 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +0000[diff] [blame]3928
Craig Topperd9021942012-08-28 07:30:47 +0000[diff] [blame]3929 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +0000[diff] [blame]3930 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
3931 Constant::getNullValue(Ty);
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3932 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3933
3934 // Create operands to load from the constant pool entry.
3935 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
3936 MOs.push_back(MachineOperand::CreateImm(1));
3937 MOs.push_back(MachineOperand::CreateReg(0, false));
3938 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
Rafael Espindola094fad32009-04-08 21:14:34 +0000[diff] [blame]3939 MOs.push_back(MachineOperand::CreateReg(0, false));
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3940 break;
3941 }
3942 default: {
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3943 // Folding a normal load. Just copy the load's address operands.
3944 unsigned NumOps = LoadMI->getDesc().getNumOperands();
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]3945 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3946 MOs.push_back(LoadMI->getOperand(i));
Dan Gohman4a0b3e12009-09-21 18:30:38 +0000[diff] [blame]3947 break;
3948 }
Dan Gohman62c939d2008-12-03 05:21:24 +0000[diff] [blame]3949 }
Evan Cheng9cef48e2009-09-11 00:39:26 +0000[diff] [blame]3950 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3951}
3952
3953
Dan Gohman8e8b8a22008-10-16 01:49:15 +0000[diff] [blame]3954bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
3955 const SmallVectorImpl<unsigned> &Ops) const {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3956 // Check switch flag
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3957 if (NoFusing) return 0;
3958
3959 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3960 switch (MI->getOpcode()) {
3961 default: return false;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3962 case X86::TEST8rr:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3963 case X86::TEST16rr:
3964 case X86::TEST32rr:
3965 case X86::TEST64rr:
3966 return true;
Jakob Stoklund Olesen60045c22011-04-30 23:00:05 +0000[diff] [blame]3967 case X86::ADD32ri:
3968 // FIXME: AsmPrinter doesn't know how to handle
3969 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3970 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3971 return false;
3972 break;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3973 }
3974 }
3975
3976 if (Ops.size() != 1)
3977 return false;
3978
3979 unsigned OpNum = Ops[0];
3980 unsigned Opc = MI->getOpcode();
Chris Lattner749c6f62008-01-07 07:27:27 +0000[diff] [blame]3981 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3982 bool isTwoAddr = NumOps > 1 &&
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]3983 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3984
3985 // Folding a memory location into the two-address part of a two-address
3986 // instruction is different than folding it other places. It requires
3987 // replacing the *two* registers with the memory location.
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3988 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]3989 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3990 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3991 } else if (OpNum == 0) { // If operand 0
3992 switch (Opc) {
Chris Lattner9ac75422009-07-14 20:19:57 +0000[diff] [blame]3993 case X86::MOV8r0:
Dan Gohmanf1b4d262010-01-12 04:42:54 +0000[diff] [blame]3994 case X86::MOV16r0:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3995 case X86::MOV32r0:
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]3996 case X86::MOV64r0: return true;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]3997 default: break;
3998 }
3999 OpcodeTablePtr = &RegOp2MemOpTable0;
4000 } else if (OpNum == 1) {
4001 OpcodeTablePtr = &RegOp2MemOpTable1;
4002 } else if (OpNum == 2) {
4003 OpcodeTablePtr = &RegOp2MemOpTable2;
Craig Topperbbdbb052012-08-31 22:12:16 +0000[diff] [blame]4004 } else if (OpNum == 3) {
4005 OpcodeTablePtr = &RegOp2MemOpTable3;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4006 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]4007
Chris Lattner99ae6652010-10-08 03:54:52 +0000[diff] [blame]4008 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4009 return true;
Jakob Stoklund Olesen1f323402010-07-09 20:43:13 +0000[diff] [blame]4010 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4011}
4012
4013bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4014 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]4015 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]4016 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4017 MemOp2RegOpTable.find(MI->getOpcode());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4018 if (I == MemOp2RegOpTable.end())
4019 return false;
4020 unsigned Opc = I->second.first;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]4021 unsigned Index = I->second.second & TB_INDEX_MASK;
4022 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4023 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4024 if (UnfoldLoad && !FoldedLoad)
4025 return false;
4026 UnfoldLoad &= FoldedLoad;
4027 if (UnfoldStore && !FoldedStore)
4028 return false;
4029 UnfoldStore &= FoldedStore;
4030
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]4031 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]4032 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng98ec91e2010-07-02 20:36:18 +0000[diff] [blame]4033 if (!MI->hasOneMemOperand() &&
4034 RC == &X86::VR128RegClass &&
4035 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4036 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4037 // conservatively assume the address is unaligned. That's bad for
4038 // performance.
4039 return false;
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]4040 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4041 SmallVector<MachineOperand,2> BeforeOps;
4042 SmallVector<MachineOperand,2> AfterOps;
4043 SmallVector<MachineOperand,4> ImpOps;
4044 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4045 MachineOperand &Op = MI->getOperand(i);
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]4046 if (i >= Index && i < Index + X86::AddrNumOperands)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4047 AddrOps.push_back(Op);
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]4048 else if (Op.isReg() && Op.isImplicit())
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4049 ImpOps.push_back(Op);
4050 else if (i < Index)
4051 BeforeOps.push_back(Op);
4052 else if (i > Index)
4053 AfterOps.push_back(Op);
4054 }
4055
4056 // Emit the load instruction.
4057 if (UnfoldLoad) {
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]4058 std::pair<MachineInstr::mmo_iterator,
4059 MachineInstr::mmo_iterator> MMOs =
4060 MF.extractLoadMemRefs(MI->memoperands_begin(),
4061 MI->memoperands_end());
4062 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4063 if (UnfoldStore) {
4064 // Address operands cannot be marked isKill.
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]4065 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4066 MachineOperand &MO = NewMIs[0]->getOperand(i);
Dan Gohmand735b802008-10-03 15:45:36 +0000[diff] [blame]4067 if (MO.isReg())
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4068 MO.setIsKill(false);
4069 }
4070 }
4071 }
4072
4073 // Emit the data processing instruction.
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]4074 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4075 MachineInstrBuilder MIB(DataMI);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]4076
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4077 if (FoldedStore)
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]4078 MIB.addReg(Reg, RegState::Define);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4079 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]4080 MIB.addOperand(BeforeOps[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4081 if (FoldedLoad)
4082 MIB.addReg(Reg);
4083 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +0000[diff] [blame]4084 MIB.addOperand(AfterOps[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4085 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4086 MachineOperand &MO = ImpOps[i];
Bill Wendling587daed2009-05-13 21:33:08 +0000[diff] [blame]4087 MIB.addReg(MO.getReg(),
4088 getDefRegState(MO.isDef()) |
4089 RegState::Implicit |
4090 getKillRegState(MO.isKill()) |
Evan Cheng4784f1f2009-06-30 08:49:04 +0000[diff] [blame]4091 getDeadRegState(MO.isDead()) |
4092 getUndefRegState(MO.isUndef()));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4093 }
4094 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4095 switch (DataMI->getOpcode()) {
4096 default: break;
4097 case X86::CMP64ri32:
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]4098 case X86::CMP64ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4099 case X86::CMP32ri:
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]4100 case X86::CMP32ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4101 case X86::CMP16ri:
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]4102 case X86::CMP16ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4103 case X86::CMP8ri: {
4104 MachineOperand &MO0 = DataMI->getOperand(0);
4105 MachineOperand &MO1 = DataMI->getOperand(1);
4106 if (MO1.getImm() == 0) {
Craig Topper195f1b82012-08-21 08:16:16 +0000[diff] [blame]4107 unsigned NewOpc;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4108 switch (DataMI->getOpcode()) {
Craig Topper195f1b82012-08-21 08:16:16 +0000[diff] [blame]4109 default: llvm_unreachable("Unreachable!");
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]4110 case X86::CMP64ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4111 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]4112 case X86::CMP32ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4113 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +0000[diff] [blame]4114 case X86::CMP16ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4115 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4116 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4117 }
Chris Lattner5080f4d2008-01-11 18:10:50 +0000[diff] [blame]4118 DataMI->setDesc(get(NewOpc));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4119 MO1.ChangeToRegister(MO0.getReg(), false);
4120 }
4121 }
4122 }
4123 NewMIs.push_back(DataMI);
4124
4125 // Emit the store instruction.
4126 if (UnfoldStore) {
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]4127 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]4128 std::pair<MachineInstr::mmo_iterator,
4129 MachineInstr::mmo_iterator> MMOs =
4130 MF.extractStoreMemRefs(MI->memoperands_begin(),
4131 MI->memoperands_end());
4132 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4133 }
4134
4135 return true;
4136}
4137
4138bool
4139X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
Bill Wendlingfbef3102009-02-11 21:51:19 +0000[diff] [blame]4140 SmallVectorImpl<SDNode*> &NewNodes) const {
Dan Gohmane8be6c62008-07-17 19:10:17 +0000[diff] [blame]4141 if (!N->isMachineOpcode())
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4142 return false;
4143
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]4144 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4145 MemOp2RegOpTable.find(N->getMachineOpcode());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4146 if (I == MemOp2RegOpTable.end())
4147 return false;
4148 unsigned Opc = I->second.first;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]4149 unsigned Index = I->second.second & TB_INDEX_MASK;
4150 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4151 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]4152 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]4153 MachineFunction &MF = DAG.getMachineFunction();
4154 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]4155 unsigned NumDefs = MCID.NumDefs;
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]4156 std::vector<SDValue> AddrOps;
4157 std::vector<SDValue> BeforeOps;
4158 std::vector<SDValue> AfterOps;
Dale Johannesened2eee62009-02-06 01:31:28 +0000[diff] [blame]4159 DebugLoc dl = N->getDebugLoc();
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4160 unsigned NumOps = N->getNumOperands();
Dan Gohmanc76909a2009-09-25 20:36:54 +0000[diff] [blame]4161 for (unsigned i = 0; i != NumOps-1; ++i) {
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]4162 SDValue Op = N->getOperand(i);
Chris Lattnerac0ed5d2010-07-08 22:41:28 +0000[diff] [blame]4163 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4164 AddrOps.push_back(Op);
Dan Gohmanb37a8202009-03-04 19:23:38 +0000[diff] [blame]4165 else if (i < Index-NumDefs)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4166 BeforeOps.push_back(Op);
Dan Gohmanb37a8202009-03-04 19:23:38 +0000[diff] [blame]4167 else if (i > Index-NumDefs)
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4168 AfterOps.push_back(Op);
4169 }
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]4170 SDValue Chain = N->getOperand(NumOps-1);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4171 AddrOps.push_back(Chain);
4172
4173 // Emit the load instruction.
4174 SDNode *Load = 0;
4175 if (FoldedLoad) {
Owen Andersone50ed302009-08-10 22:56:29 +0000[diff] [blame]4176 EVT VT = *RC->vt_begin();
Evan Cheng600c0432009-11-16 21:56:03 +0000[diff] [blame]4177 std::pair<MachineInstr::mmo_iterator,
4178 MachineInstr::mmo_iterator> MMOs =
4179 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4180 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng98ec91e2010-07-02 20:36:18 +0000[diff] [blame]4181 if (!(*MMOs.first) &&
4182 RC == &X86::VR128RegClass &&
4183 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4184 // Do not introduce a slow unaligned load.
4185 return false;
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]4186 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4187 bool isAligned = (*MMOs.first) &&
4188 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman602b0c82009-09-25 18:54:59 +0000[diff] [blame]4189 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
4190 VT, MVT::Other, &AddrOps[0], AddrOps.size());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4191 NewNodes.push_back(Load);
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]4192
4193 // Preserve memory reference information.
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]4194 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4195 }
4196
4197 // Emit the data processing instruction.
Owen Andersone50ed302009-08-10 22:56:29 +0000[diff] [blame]4198 std::vector<EVT> VTs;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4199 const TargetRegisterClass *DstRC = 0;
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]4200 if (MCID.getNumDefs() > 0) {
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +0000[diff] [blame]4201 DstRC = getRegClass(MCID, 0, &RI, MF);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4202 VTs.push_back(*DstRC->vt_begin());
4203 }
4204 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
Owen Andersone50ed302009-08-10 22:56:29 +0000[diff] [blame]4205 EVT VT = N->getValueType(i);
Evan Chenge837dea2011-06-28 19:10:37 +0000[diff] [blame]4206 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4207 VTs.push_back(VT);
4208 }
4209 if (Load)
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]4210 BeforeOps.push_back(SDValue(Load, 0));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4211 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
Dan Gohman602b0c82009-09-25 18:54:59 +0000[diff] [blame]4212 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0],
4213 BeforeOps.size());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4214 NewNodes.push_back(NewNode);
4215
4216 // Emit the store instruction.
4217 if (FoldedStore) {
4218 AddrOps.pop_back();
Dan Gohman475871a2008-07-27 21:46:04 +0000[diff] [blame]4219 AddrOps.push_back(SDValue(NewNode, 0));
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4220 AddrOps.push_back(Chain);
Evan Cheng600c0432009-11-16 21:56:03 +0000[diff] [blame]4221 std::pair<MachineInstr::mmo_iterator,
4222 MachineInstr::mmo_iterator> MMOs =
4223 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4224 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng98ec91e2010-07-02 20:36:18 +0000[diff] [blame]4225 if (!(*MMOs.first) &&
4226 RC == &X86::VR128RegClass &&
4227 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4228 // Do not introduce a slow unaligned store.
4229 return false;
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]4230 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4231 bool isAligned = (*MMOs.first) &&
4232 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman602b0c82009-09-25 18:54:59 +0000[diff] [blame]4233 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4234 isAligned, TM),
4235 dl, MVT::Other,
4236 &AddrOps[0], AddrOps.size());
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4237 NewNodes.push_back(Store);
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]4238
4239 // Preserve memory reference information.
Dan Gohman91e69c32009-10-09 18:10:05 +0000[diff] [blame]4240 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4241 }
4242
4243 return true;
4244}
4245
4246unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
Dan Gohman0115e162009-10-30 22:18:41 +0000[diff] [blame]4247 bool UnfoldLoad, bool UnfoldStore,
4248 unsigned *LoadRegIndex) const {
Chris Lattner45a1cb22010-10-07 23:08:41 +0000[diff] [blame]4249 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4250 MemOp2RegOpTable.find(Opc);
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4251 if (I == MemOp2RegOpTable.end())
4252 return 0;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]4253 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4254 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4255 if (UnfoldLoad && !FoldedLoad)
4256 return 0;
4257 if (UnfoldStore && !FoldedStore)
4258 return 0;
Dan Gohman0115e162009-10-30 22:18:41 +0000[diff] [blame]4259 if (LoadRegIndex)
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000[diff] [blame]4260 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
Owen Anderson43dbe052008-01-07 01:35:02 +0000[diff] [blame]4261 return I->second.first;
4262}
4263
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4264bool
4265X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4266 int64_t &Offset1, int64_t &Offset2) const {
4267 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4268 return false;
4269 unsigned Opc1 = Load1->getMachineOpcode();
4270 unsigned Opc2 = Load2->getMachineOpcode();
4271 switch (Opc1) {
4272 default: return false;
4273 case X86::MOV8rm:
4274 case X86::MOV16rm:
4275 case X86::MOV32rm:
4276 case X86::MOV64rm:
4277 case X86::LD_Fp32m:
4278 case X86::LD_Fp64m:
4279 case X86::LD_Fp80m:
4280 case X86::MOVSSrm:
4281 case X86::MOVSDrm:
4282 case X86::MMX_MOVD64rm:
4283 case X86::MMX_MOVQ64rm:
4284 case X86::FsMOVAPSrm:
4285 case X86::FsMOVAPDrm:
4286 case X86::MOVAPSrm:
4287 case X86::MOVUPSrm:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4288 case X86::MOVAPDrm:
4289 case X86::MOVDQArm:
4290 case X86::MOVDQUrm:
Bruno Cardoso Lopesb4e905d2011-09-15 22:15:52 +0000[diff] [blame]4291 // AVX load instructions
4292 case X86::VMOVSSrm:
4293 case X86::VMOVSDrm:
4294 case X86::FsVMOVAPSrm:
4295 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]4296 case X86::VMOVAPSrm:
4297 case X86::VMOVUPSrm:
4298 case X86::VMOVAPDrm:
4299 case X86::VMOVDQArm:
4300 case X86::VMOVDQUrm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]4301 case X86::VMOVAPSYrm:
4302 case X86::VMOVUPSYrm:
4303 case X86::VMOVAPDYrm:
4304 case X86::VMOVDQAYrm:
4305 case X86::VMOVDQUYrm:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4306 break;
4307 }
4308 switch (Opc2) {
4309 default: return false;
4310 case X86::MOV8rm:
4311 case X86::MOV16rm:
4312 case X86::MOV32rm:
4313 case X86::MOV64rm:
4314 case X86::LD_Fp32m:
4315 case X86::LD_Fp64m:
4316 case X86::LD_Fp80m:
4317 case X86::MOVSSrm:
4318 case X86::MOVSDrm:
4319 case X86::MMX_MOVD64rm:
4320 case X86::MMX_MOVQ64rm:
4321 case X86::FsMOVAPSrm:
4322 case X86::FsMOVAPDrm:
4323 case X86::MOVAPSrm:
4324 case X86::MOVUPSrm:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4325 case X86::MOVAPDrm:
4326 case X86::MOVDQArm:
4327 case X86::MOVDQUrm:
Bruno Cardoso Lopesb4e905d2011-09-15 22:15:52 +0000[diff] [blame]4328 // AVX load instructions
4329 case X86::VMOVSSrm:
4330 case X86::VMOVSDrm:
4331 case X86::FsVMOVAPSrm:
4332 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000[diff] [blame]4333 case X86::VMOVAPSrm:
4334 case X86::VMOVUPSrm:
4335 case X86::VMOVAPDrm:
4336 case X86::VMOVDQArm:
4337 case X86::VMOVDQUrm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]4338 case X86::VMOVAPSYrm:
4339 case X86::VMOVUPSYrm:
4340 case X86::VMOVAPDYrm:
4341 case X86::VMOVDQAYrm:
4342 case X86::VMOVDQUYrm:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4343 break;
4344 }
4345
4346 // Check if chain operands and base addresses match.
4347 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4348 Load1->getOperand(5) != Load2->getOperand(5))
4349 return false;
4350 // Segment operands should match as well.
4351 if (Load1->getOperand(4) != Load2->getOperand(4))
4352 return false;
4353 // Scale should be 1, Index should be Reg0.
4354 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4355 Load1->getOperand(2) == Load2->getOperand(2)) {
4356 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4357 return false;
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4358
4359 // Now let's examine the displacements.
4360 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4361 isa<ConstantSDNode>(Load2->getOperand(3))) {
4362 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4363 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4364 return true;
4365 }
4366 }
4367 return false;
4368}
4369
4370bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4371 int64_t Offset1, int64_t Offset2,
4372 unsigned NumLoads) const {
4373 assert(Offset2 > Offset1);
4374 if ((Offset2 - Offset1) / 8 > 64)
4375 return false;
4376
4377 unsigned Opc1 = Load1->getMachineOpcode();
4378 unsigned Opc2 = Load2->getMachineOpcode();
4379 if (Opc1 != Opc2)
4380 return false; // FIXME: overly conservative?
4381
4382 switch (Opc1) {
4383 default: break;
4384 case X86::LD_Fp32m:
4385 case X86::LD_Fp64m:
4386 case X86::LD_Fp80m:
4387 case X86::MMX_MOVD64rm:
4388 case X86::MMX_MOVQ64rm:
4389 return false;
4390 }
4391
4392 EVT VT = Load1->getValueType(0);
4393 switch (VT.getSimpleVT().SimpleTy) {
Bill Wendling19d85972010-06-22 22:16:17 +0000[diff] [blame]4394 default:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4395 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4396 // have 16 of them to play with.
4397 if (TM.getSubtargetImpl()->is64Bit()) {
4398 if (NumLoads >= 3)
4399 return false;
Bill Wendling19d85972010-06-22 22:16:17 +0000[diff] [blame]4400 } else if (NumLoads) {
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4401 return false;
Bill Wendling19d85972010-06-22 22:16:17 +0000[diff] [blame]4402 }
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4403 break;
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4404 case MVT::i8:
4405 case MVT::i16:
4406 case MVT::i32:
4407 case MVT::i64:
Evan Chengafc36732010-01-22 23:49:11 +0000[diff] [blame]4408 case MVT::f32:
4409 case MVT::f64:
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4410 if (NumLoads)
4411 return false;
Bill Wendling19d85972010-06-22 22:16:17 +0000[diff] [blame]4412 break;
Evan Cheng96dc1152010-01-22 03:34:51 +0000[diff] [blame]4413 }
4414
4415 return true;
4416}
4417
4418
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]4419bool X86InstrInfo::
Owen Anderson44eb65c2008-08-14 22:49:33 +0000[diff] [blame]4420ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]4421 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
Evan Cheng97af60b2008-08-29 23:21:31 +0000[diff] [blame]4422 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
Dan Gohman279c22e2008-10-21 03:29:32 +0000[diff] [blame]4423 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4424 return true;
Evan Cheng97af60b2008-08-29 23:21:31 +0000[diff] [blame]4425 Cond[0].setImm(GetOppositeBranchCondition(CC));
Chris Lattner9cd68752006-10-21 05:52:40 +0000[diff] [blame]4426 return false;
Chris Lattner7fbe9722006-10-20 17:42:20 +0000[diff] [blame]4427}
4428
Evan Cheng23066282008-10-27 07:14:50 +0000[diff] [blame]4429bool X86InstrInfo::
Evan Cheng4350eb82009-02-06 17:17:30 +0000[diff] [blame]4430isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4431 // FIXME: Return false for x87 stack register classes for now. We can't
Evan Cheng23066282008-10-27 07:14:50 +0000[diff] [blame]4432 // allow any loads of these registers before FpGet_ST0_80.
Evan Cheng4350eb82009-02-06 17:17:30 +0000[diff] [blame]4433 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4434 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
Evan Cheng23066282008-10-27 07:14:50 +0000[diff] [blame]4435}
4436
Dan Gohman57c3dac2008-09-30 00:58:23 +0000[diff] [blame]4437/// getGlobalBaseReg - Return a virtual register initialized with the
4438/// the global base register value. Output instructions required to
4439/// initialize the register in the function entry block, if necessary.
Dan Gohman8b746962008-09-23 18:22:58 +0000[diff] [blame]4440///
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4441/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4442///
Dan Gohman57c3dac2008-09-30 00:58:23 +0000[diff] [blame]4443unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4444 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4445 "X86-64 PIC uses RIP relative addressing");
4446
4447 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4448 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4449 if (GlobalBaseReg != 0)
4450 return GlobalBaseReg;
4451
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4452 // Create the register. The code to initialize it is inserted
4453 // later, by the CGBR pass (below).
Dan Gohman8b746962008-09-23 18:22:58 +0000[diff] [blame]4454 MachineRegisterInfo &RegInfo = MF->getRegInfo();
Jakob Stoklund Olesen53df9252012-05-20 18:43:00 +0000[diff] [blame]4455 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Dan Gohman57c3dac2008-09-30 00:58:23 +0000[diff] [blame]4456 X86FI->setGlobalBaseReg(GlobalBaseReg);
4457 return GlobalBaseReg;
Dan Gohman8b746962008-09-23 18:22:58 +0000[diff] [blame]4458}
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4459
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4460// These are the replaceable SSE instructions. Some of these have Int variants
4461// that we don't include here. We don't want to replace instructions selected
4462// by intrinsics.
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4463static const uint16_t ReplaceableInstrs[][3] = {
Bruno Cardoso Lopes4d043622010-08-12 02:08:52 +0000[diff] [blame]4464 //PackedSingle PackedDouble PackedInt
Jakob Stoklund Olesen357be7f2010-03-30 22:46:53 +0000[diff] [blame]4465 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4466 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4467 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4468 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4469 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4470 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4471 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4472 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4473 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4474 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4475 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4476 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4477 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4478 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]4479 // AVX 128-bit support
4480 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4481 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4482 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4483 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4484 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4485 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4486 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4487 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4488 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4489 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4490 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4491 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +0000[diff] [blame]4492 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4493 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +0000[diff] [blame]4494 // AVX 256-bit support
4495 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4496 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4497 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4498 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4499 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
Craig Topper4c077a12011-11-15 05:55:35 +0000[diff] [blame]4500 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4501};
4502
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4503static const uint16_t ReplaceableInstrsAVX2[][3] = {
Craig Topper4c077a12011-11-15 05:55:35 +0000[diff] [blame]4504 //PackedSingle PackedDouble PackedInt
Craig Topperb80ada92011-11-09 09:37:21 +0000[diff] [blame]4505 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4506 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4507 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4508 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4509 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4510 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4511 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
Craig Topperfe2a6c52011-11-29 05:37:58 +0000[diff] [blame]4512 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4513 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4514 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4515 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4516 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4517 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4518 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4519};
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4520
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4521// FIXME: Some shuffle and unpack instructions have equivalents in different
4522// domains, but they require a bit more work than just switching opcodes.
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4523
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4524static const uint16_t *lookup(unsigned opcode, unsigned domain) {
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4525 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4526 if (ReplaceableInstrs[i][domain-1] == opcode)
4527 return ReplaceableInstrs[i];
Craig Topper44ec9fd2011-11-15 06:39:01 +0000[diff] [blame]4528 return 0;
4529}
4530
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4531static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
Craig Topper44ec9fd2011-11-15 06:39:01 +0000[diff] [blame]4532 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4533 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4534 return ReplaceableInstrsAVX2[i];
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4535 return 0;
4536}
4537
4538std::pair<uint16_t, uint16_t>
Jakob Stoklund Olesen98e933f2011-09-27 22:57:18 +0000[diff] [blame]4539X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4540 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
Craig Topper4c077a12011-11-15 05:55:35 +0000[diff] [blame]4541 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
Craig Topper44ec9fd2011-11-15 06:39:01 +0000[diff] [blame]4542 uint16_t validDomains = 0;
4543 if (domain && lookup(MI->getOpcode(), domain))
4544 validDomains = 0xe;
4545 else if (domain && lookupAVX2(MI->getOpcode(), domain))
4546 validDomains = hasAVX2 ? 0xe : 0x6;
4547 return std::make_pair(domain, validDomains);
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4548}
4549
Jakob Stoklund Olesen98e933f2011-09-27 22:57:18 +0000[diff] [blame]4550void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4551 assert(Domain>0 && Domain<4 && "Invalid execution domain");
4552 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4553 assert(dom && "Not an SSE instruction");
Craig Topper72051bf2012-03-09 07:45:21 +0000[diff] [blame]4554 const uint16_t *table = lookup(MI->getOpcode(), dom);
Jakob Stoklund Olesen7f5e43f2011-11-23 04:03:08 +0000[diff] [blame]4555 if (!table) { // try the other table
4556 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
4557 "256-bit vector operations only available in AVX2");
Craig Topper44ec9fd2011-11-15 06:39:01 +0000[diff] [blame]4558 table = lookupAVX2(MI->getOpcode(), dom);
Jakob Stoklund Olesen7f5e43f2011-11-23 04:03:08 +0000[diff] [blame]4559 }
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +0000[diff] [blame]4560 assert(table && "Cannot change domain");
4561 MI->setDesc(get(table[Domain-1]));
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +0000[diff] [blame]4562}
Chris Lattneree9eb412010-04-26 23:37:21 +0000[diff] [blame]4563
4564/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
4565void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
4566 NopInst.setOpcode(X86::NOOP);
4567}
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4568
Andrew Tricke0ef5092011-03-05 08:00:22 +0000[diff] [blame]4569bool X86InstrInfo::isHighLatencyDef(int opc) const {
4570 switch (opc) {
Evan Cheng23128422010-10-19 18:58:51 +0000[diff] [blame]4571 default: return false;
4572 case X86::DIVSDrm:
4573 case X86::DIVSDrm_Int:
4574 case X86::DIVSDrr:
4575 case X86::DIVSDrr_Int:
4576 case X86::DIVSSrm:
4577 case X86::DIVSSrm_Int:
4578 case X86::DIVSSrr:
4579 case X86::DIVSSrr_Int:
4580 case X86::SQRTPDm:
4581 case X86::SQRTPDm_Int:
4582 case X86::SQRTPDr:
4583 case X86::SQRTPDr_Int:
4584 case X86::SQRTPSm:
4585 case X86::SQRTPSm_Int:
4586 case X86::SQRTPSr:
4587 case X86::SQRTPSr_Int:
4588 case X86::SQRTSDm:
4589 case X86::SQRTSDm_Int:
4590 case X86::SQRTSDr:
4591 case X86::SQRTSDr_Int:
4592 case X86::SQRTSSm:
4593 case X86::SQRTSSm_Int:
4594 case X86::SQRTSSr:
4595 case X86::SQRTSSr_Int:
Bruno Cardoso Lopesb4e905d2011-09-15 22:15:52 +0000[diff] [blame]4596 // AVX instructions with high latency
4597 case X86::VDIVSDrm:
4598 case X86::VDIVSDrm_Int:
4599 case X86::VDIVSDrr:
4600 case X86::VDIVSDrr_Int:
4601 case X86::VDIVSSrm:
4602 case X86::VDIVSSrm_Int:
4603 case X86::VDIVSSrr:
4604 case X86::VDIVSSrr_Int:
4605 case X86::VSQRTPDm:
4606 case X86::VSQRTPDm_Int:
4607 case X86::VSQRTPDr:
4608 case X86::VSQRTPDr_Int:
4609 case X86::VSQRTPSm:
4610 case X86::VSQRTPSm_Int:
4611 case X86::VSQRTPSr:
4612 case X86::VSQRTPSr_Int:
4613 case X86::VSQRTSDm:
4614 case X86::VSQRTSDm_Int:
4615 case X86::VSQRTSDr:
4616 case X86::VSQRTSSm:
4617 case X86::VSQRTSSm_Int:
4618 case X86::VSQRTSSr:
Evan Cheng23128422010-10-19 18:58:51 +0000[diff] [blame]4619 return true;
4620 }
4621}
4622
Andrew Tricke0ef5092011-03-05 08:00:22 +0000[diff] [blame]4623bool X86InstrInfo::
4624hasHighOperandLatency(const InstrItineraryData *ItinData,
4625 const MachineRegisterInfo *MRI,
4626 const MachineInstr *DefMI, unsigned DefIdx,
4627 const MachineInstr *UseMI, unsigned UseIdx) const {
4628 return isHighLatencyDef(DefMI->getOpcode());
4629}
4630
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4631namespace {
4632 /// CGBR - Create Global Base Reg pass. This initializes the PIC
4633 /// global base register for x86-32.
4634 struct CGBR : public MachineFunctionPass {
4635 static char ID;
Owen Anderson90c579d2010-08-06 18:33:48 +0000[diff] [blame]4636 CGBR() : MachineFunctionPass(ID) {}
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4637
4638 virtual bool runOnMachineFunction(MachineFunction &MF) {
4639 const X86TargetMachine *TM =
4640 static_cast<const X86TargetMachine *>(&MF.getTarget());
4641
4642 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
4643 "X86-64 PIC uses RIP relative addressing");
4644
4645 // Only emit a global base reg in PIC mode.
4646 if (TM->getRelocationModel() != Reloc::PIC_)
4647 return false;
4648
Dan Gohmand8c0a512010-09-17 20:24:24 +0000[diff] [blame]4649 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
4650 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4651
4652 // If we didn't need a GlobalBaseReg, don't insert code.
4653 if (GlobalBaseReg == 0)
4654 return false;
4655
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4656 // Insert the set of GlobalBaseReg into the first MBB of the function
4657 MachineBasicBlock &FirstMBB = MF.front();
4658 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
4659 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
4660 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4661 const X86InstrInfo *TII = TM->getInstrInfo();
4662
4663 unsigned PC;
4664 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
Craig Topperc9099502012-04-20 06:31:50 +0000[diff] [blame]4665 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4666 else
Dan Gohmand8c0a512010-09-17 20:24:24 +0000[diff] [blame]4667 PC = GlobalBaseReg;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]4668
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4669 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4670 // only used in JIT code emission as displacement to pc.
4671 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000[diff] [blame]4672
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4673 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4674 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4675 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
Dan Gohman84023e02010-07-10 09:00:22 +0000[diff] [blame]4676 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4677 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4678 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4679 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4680 }
4681
4682 return true;
4683 }
4684
4685 virtual const char *getPassName() const {
4686 return "X86 PIC Global Base Reg Initialization";
4687 }
4688
4689 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4690 AU.setPreservesCFG();
4691 MachineFunctionPass::getAnalysisUsage(AU);
4692 }
4693 };
4694}
4695
4696char CGBR::ID = 0;
4697FunctionPass*
4698llvm::createGlobalBaseRegPass() { return new CGBR(); }
Hans Wennborgf0234fc2012-06-01 16:27:21 +0000[diff] [blame]4699
4700namespace {
4701 struct LDTLSCleanup : public MachineFunctionPass {
4702 static char ID;
4703 LDTLSCleanup() : MachineFunctionPass(ID) {}
4704
4705 virtual bool runOnMachineFunction(MachineFunction &MF) {
4706 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4707 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4708 // No point folding accesses if there isn't at least two.
4709 return false;
4710 }
4711
4712 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4713 return VisitNode(DT->getRootNode(), 0);
4714 }
4715
4716 // Visit the dominator subtree rooted at Node in pre-order.
4717 // If TLSBaseAddrReg is non-null, then use that to replace any
4718 // TLS_base_addr instructions. Otherwise, create the register
4719 // when the first such instruction is seen, and then use it
4720 // as we encounter more instructions.
4721 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4722 MachineBasicBlock *BB = Node->getBlock();
4723 bool Changed = false;
4724
4725 // Traverse the current block.
4726 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4727 ++I) {
4728 switch (I->getOpcode()) {
4729 case X86::TLS_base_addr32:
4730 case X86::TLS_base_addr64:
4731 if (TLSBaseAddrReg)
4732 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4733 else
4734 I = SetRegister(I, &TLSBaseAddrReg);
4735 Changed = true;
4736 break;
4737 default:
4738 break;
4739 }
4740 }
4741
4742 // Visit the children of this block in the dominator tree.
4743 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4744 I != E; ++I) {
4745 Changed |= VisitNode(*I, TLSBaseAddrReg);
4746 }
4747
4748 return Changed;
4749 }
4750
4751 // Replace the TLS_base_addr instruction I with a copy from
4752 // TLSBaseAddrReg, returning the new instruction.
4753 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4754 unsigned TLSBaseAddrReg) {
4755 MachineFunction *MF = I->getParent()->getParent();
4756 const X86TargetMachine *TM =
4757 static_cast<const X86TargetMachine *>(&MF->getTarget());
4758 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4759 const X86InstrInfo *TII = TM->getInstrInfo();
4760
4761 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4762 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4763 TII->get(TargetOpcode::COPY),
4764 is64Bit ? X86::RAX : X86::EAX)
4765 .addReg(TLSBaseAddrReg);
4766
4767 // Erase the TLS_base_addr instruction.
4768 I->eraseFromParent();
4769
4770 return Copy;
4771 }
4772
4773 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4774 // inserting a copy instruction after I. Returns the new instruction.
4775 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4776 MachineFunction *MF = I->getParent()->getParent();
4777 const X86TargetMachine *TM =
4778 static_cast<const X86TargetMachine *>(&MF->getTarget());
4779 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4780 const X86InstrInfo *TII = TM->getInstrInfo();
4781
4782 // Create a virtual register for the TLS base address.
4783 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4784 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4785 ? &X86::GR64RegClass
4786 : &X86::GR32RegClass);
4787
4788 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4789 MachineInstr *Next = I->getNextNode();
4790 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4791 TII->get(TargetOpcode::COPY),
4792 *TLSBaseAddrReg)
4793 .addReg(is64Bit ? X86::RAX : X86::EAX);
4794
4795 return Copy;
4796 }
4797
4798 virtual const char *getPassName() const {
4799 return "Local Dynamic TLS Access Clean-up";
4800 }
4801
4802 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4803 AU.setPreservesCFG();
4804 AU.addRequired<MachineDominatorTree>();
4805 MachineFunctionPass::getAnalysisUsage(AU);
4806 }
4807 };
4808}
4809
4810char LDTLSCleanup::ID = 0;
4811FunctionPass*
4812llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }