blob: c78c24bf47b71019e09b871e957c93ac6cb39fa3 [file] [log] [blame]
Akira Hatanaka44ebe002013-03-14 19:09:52 +00001//===-- MipsSEISelLowering.cpp - MipsSE DAG Lowering Interface --*- C++ -*-===//
Akira Hatanaka96ca1822013-03-13 00:54:29 +00002//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// Subclass of MipsTargetLowering specialized for mips32/64.
11//
12//===----------------------------------------------------------------------===//
13#include "MipsSEISelLowering.h"
Eric Christopher79cc1e32014-09-02 22:28:02 +000014#include "MipsMachineFunction.h"
Akira Hatanaka96ca1822013-03-13 00:54:29 +000015#include "MipsRegisterInfo.h"
16#include "MipsTargetMachine.h"
Simon Dardis548a53f2017-01-10 16:40:57 +000017#include "llvm/ADT/APInt.h"
Akira Hatanaka96ca1822013-03-13 00:54:29 +000018#include "llvm/CodeGen/MachineInstrBuilder.h"
19#include "llvm/CodeGen/MachineRegisterInfo.h"
Akira Hatanakaa6bbde52013-04-13 02:13:30 +000020#include "llvm/IR/Intrinsics.h"
Akira Hatanaka96ca1822013-03-13 00:54:29 +000021#include "llvm/Support/CommandLine.h"
Daniel Sanders62aeab82013-10-30 13:31:27 +000022#include "llvm/Support/Debug.h"
Simon Dardis548a53f2017-01-10 16:40:57 +000023#include "llvm/Support/ErrorHandling.h"
Hans Wennborg3e9b1c12013-10-30 16:10:10 +000024#include "llvm/Support/raw_ostream.h"
Akira Hatanaka96ca1822013-03-13 00:54:29 +000025#include "llvm/Target/TargetInstrInfo.h"
26
27using namespace llvm;
28
Chandler Carruth84e68b22014-04-22 02:41:26 +000029#define DEBUG_TYPE "mips-isel"
30
Akira Hatanaka96ca1822013-03-13 00:54:29 +000031static cl::opt<bool>
Simon Dardis57f4ae42016-08-04 09:17:07 +000032UseMipsTailCalls("mips-tail-calls", cl::Hidden,
Simon Dardisd2ed8ab2016-09-27 13:15:54 +000033 cl::desc("MIPS: permit tail calls."), cl::init(false));
Akira Hatanaka96ca1822013-03-13 00:54:29 +000034
Akira Hatanaka63791212013-09-07 00:52:30 +000035static cl::opt<bool> NoDPLoadStore("mno-ldc1-sdc1", cl::init(false),
36 cl::desc("Expand double precision loads and "
37 "stores to their single precision "
38 "counterparts"));
39
Eric Christopherb1526602014-09-19 23:30:42 +000040MipsSETargetLowering::MipsSETargetLowering(const MipsTargetMachine &TM,
Eric Christopher8924d272014-07-18 23:25:04 +000041 const MipsSubtarget &STI)
42 : MipsTargetLowering(TM, STI) {
Akira Hatanaka96ca1822013-03-13 00:54:29 +000043 // Set up the register classes
Akira Hatanaka13e6ccf2013-08-06 23:08:38 +000044 addRegisterClass(MVT::i32, &Mips::GPR32RegClass);
Akira Hatanaka96ca1822013-03-13 00:54:29 +000045
Eric Christopher1c29a652014-07-18 22:55:25 +000046 if (Subtarget.isGP64bit())
Akira Hatanaka13e6ccf2013-08-06 23:08:38 +000047 addRegisterClass(MVT::i64, &Mips::GPR64RegClass);
Akira Hatanaka96ca1822013-03-13 00:54:29 +000048
Eric Christopher1c29a652014-07-18 22:55:25 +000049 if (Subtarget.hasDSP() || Subtarget.hasMSA()) {
Daniel Sanders36c671e2013-09-27 09:44:59 +000050 // Expand all truncating stores and extending loads.
Ahmed Bougacha67dd2d22015-01-07 21:27:10 +000051 for (MVT VT0 : MVT::vector_valuetypes()) {
Ahmed Bougacha2b6917b2015-01-08 00:51:32 +000052 for (MVT VT1 : MVT::vector_valuetypes()) {
Ahmed Bougacha67dd2d22015-01-07 21:27:10 +000053 setTruncStoreAction(VT0, VT1, Expand);
Ahmed Bougacha2b6917b2015-01-08 00:51:32 +000054 setLoadExtAction(ISD::SEXTLOAD, VT0, VT1, Expand);
55 setLoadExtAction(ISD::ZEXTLOAD, VT0, VT1, Expand);
56 setLoadExtAction(ISD::EXTLOAD, VT0, VT1, Expand);
57 }
Daniel Sanders36c671e2013-09-27 09:44:59 +000058 }
59 }
60
Eric Christopher1c29a652014-07-18 22:55:25 +000061 if (Subtarget.hasDSP()) {
Akira Hatanaka96ca1822013-03-13 00:54:29 +000062 MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8};
63
64 for (unsigned i = 0; i < array_lengthof(VecTys); ++i) {
Akira Hatanaka654655f2013-08-14 00:53:38 +000065 addRegisterClass(VecTys[i], &Mips::DSPRRegClass);
Akira Hatanaka96ca1822013-03-13 00:54:29 +000066
67 // Expand all builtin opcodes.
68 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
69 setOperationAction(Opc, VecTys[i], Expand);
70
Akira Hatanaka2f088222013-04-13 00:55:41 +000071 setOperationAction(ISD::ADD, VecTys[i], Legal);
72 setOperationAction(ISD::SUB, VecTys[i], Legal);
Akira Hatanaka96ca1822013-03-13 00:54:29 +000073 setOperationAction(ISD::LOAD, VecTys[i], Legal);
74 setOperationAction(ISD::STORE, VecTys[i], Legal);
75 setOperationAction(ISD::BITCAST, VecTys[i], Legal);
76 }
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +000077
78 setTargetDAGCombine(ISD::SHL);
79 setTargetDAGCombine(ISD::SRA);
80 setTargetDAGCombine(ISD::SRL);
Akira Hatanaka68741cc2013-04-30 22:37:26 +000081 setTargetDAGCombine(ISD::SETCC);
82 setTargetDAGCombine(ISD::VSELECT);
Akira Hatanaka96ca1822013-03-13 00:54:29 +000083 }
84
Eric Christopher1c29a652014-07-18 22:55:25 +000085 if (Subtarget.hasDSPR2())
Akira Hatanaka2f088222013-04-13 00:55:41 +000086 setOperationAction(ISD::MUL, MVT::v2i16, Legal);
87
Eric Christopher1c29a652014-07-18 22:55:25 +000088 if (Subtarget.hasMSA()) {
Daniel Sandersc65f58a2013-09-11 10:15:48 +000089 addMSAIntType(MVT::v16i8, &Mips::MSA128BRegClass);
90 addMSAIntType(MVT::v8i16, &Mips::MSA128HRegClass);
91 addMSAIntType(MVT::v4i32, &Mips::MSA128WRegClass);
92 addMSAIntType(MVT::v2i64, &Mips::MSA128DRegClass);
93 addMSAFloatType(MVT::v8f16, &Mips::MSA128HRegClass);
94 addMSAFloatType(MVT::v4f32, &Mips::MSA128WRegClass);
95 addMSAFloatType(MVT::v2f64, &Mips::MSA128DRegClass);
Daniel Sandersf7456c72013-09-23 13:22:24 +000096
Simon Dardis0e2ee3b2016-11-18 16:17:44 +000097 // f16 is a storage-only type, always promote it to f32.
98 addRegisterClass(MVT::f16, &Mips::MSA128HRegClass);
99 setOperationAction(ISD::SETCC, MVT::f16, Promote);
100 setOperationAction(ISD::BR_CC, MVT::f16, Promote);
101 setOperationAction(ISD::SELECT_CC, MVT::f16, Promote);
102 setOperationAction(ISD::SELECT, MVT::f16, Promote);
103 setOperationAction(ISD::FADD, MVT::f16, Promote);
104 setOperationAction(ISD::FSUB, MVT::f16, Promote);
105 setOperationAction(ISD::FMUL, MVT::f16, Promote);
106 setOperationAction(ISD::FDIV, MVT::f16, Promote);
107 setOperationAction(ISD::FREM, MVT::f16, Promote);
108 setOperationAction(ISD::FMA, MVT::f16, Promote);
109 setOperationAction(ISD::FNEG, MVT::f16, Promote);
110 setOperationAction(ISD::FABS, MVT::f16, Promote);
111 setOperationAction(ISD::FCEIL, MVT::f16, Promote);
112 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
113 setOperationAction(ISD::FCOS, MVT::f16, Promote);
114 setOperationAction(ISD::FP_EXTEND, MVT::f16, Promote);
115 setOperationAction(ISD::FFLOOR, MVT::f16, Promote);
116 setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
117 setOperationAction(ISD::FPOW, MVT::f16, Promote);
118 setOperationAction(ISD::FPOWI, MVT::f16, Promote);
119 setOperationAction(ISD::FRINT, MVT::f16, Promote);
120 setOperationAction(ISD::FSIN, MVT::f16, Promote);
121 setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
122 setOperationAction(ISD::FSQRT, MVT::f16, Promote);
123 setOperationAction(ISD::FEXP, MVT::f16, Promote);
124 setOperationAction(ISD::FEXP2, MVT::f16, Promote);
125 setOperationAction(ISD::FLOG, MVT::f16, Promote);
126 setOperationAction(ISD::FLOG2, MVT::f16, Promote);
127 setOperationAction(ISD::FLOG10, MVT::f16, Promote);
128 setOperationAction(ISD::FROUND, MVT::f16, Promote);
129 setOperationAction(ISD::FTRUNC, MVT::f16, Promote);
130 setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
131 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
132 setOperationAction(ISD::FMINNAN, MVT::f16, Promote);
133 setOperationAction(ISD::FMAXNAN, MVT::f16, Promote);
134
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000135 setTargetDAGCombine(ISD::AND);
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000136 setTargetDAGCombine(ISD::OR);
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000137 setTargetDAGCombine(ISD::SRA);
Daniel Sanderse1d24352013-09-24 12:04:44 +0000138 setTargetDAGCombine(ISD::VSELECT);
Daniel Sandersf7456c72013-09-23 13:22:24 +0000139 setTargetDAGCombine(ISD::XOR);
Jack Carter3a2c2d42013-08-13 20:54:07 +0000140 }
141
Eric Christophere8ae3e32015-05-07 23:10:21 +0000142 if (!Subtarget.useSoftFloat()) {
Akira Hatanaka96ca1822013-03-13 00:54:29 +0000143 addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
144
145 // When dealing with single precision only, use libcalls
Eric Christopher1c29a652014-07-18 22:55:25 +0000146 if (!Subtarget.isSingleFloat()) {
147 if (Subtarget.isFP64bit())
Akira Hatanaka96ca1822013-03-13 00:54:29 +0000148 addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
149 else
150 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
151 }
152 }
153
Akira Hatanakabe8612f2013-03-30 01:36:35 +0000154 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom);
155 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom);
156 setOperationAction(ISD::MULHS, MVT::i32, Custom);
157 setOperationAction(ISD::MULHU, MVT::i32, Custom);
158
Eric Christopher1c29a652014-07-18 22:55:25 +0000159 if (Subtarget.hasCnMips())
Kai Nacke93fe5e82014-03-20 11:51:58 +0000160 setOperationAction(ISD::MUL, MVT::i64, Legal);
Eric Christopher1c29a652014-07-18 22:55:25 +0000161 else if (Subtarget.isGP64bit())
Kai Nacke93fe5e82014-03-20 11:51:58 +0000162 setOperationAction(ISD::MUL, MVT::i64, Custom);
163
Eric Christopher1c29a652014-07-18 22:55:25 +0000164 if (Subtarget.isGP64bit()) {
Vasileios Kalintirisef96a8e2015-01-26 12:33:22 +0000165 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Custom);
166 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Custom);
Akira Hatanaka4f1130e2013-04-11 19:29:26 +0000167 setOperationAction(ISD::MULHS, MVT::i64, Custom);
168 setOperationAction(ISD::MULHU, MVT::i64, Custom);
Jan Vesely54468a5a2014-10-17 14:45:28 +0000169 setOperationAction(ISD::SDIVREM, MVT::i64, Custom);
170 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
Akira Hatanaka4f1130e2013-04-11 19:29:26 +0000171 }
Akira Hatanakabe8612f2013-03-30 01:36:35 +0000172
Akira Hatanakaa6bbde52013-04-13 02:13:30 +0000173 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
174 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
175
Akira Hatanakabe8612f2013-03-30 01:36:35 +0000176 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
177 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
Akira Hatanaka96ca1822013-03-13 00:54:29 +0000178 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
179 setOperationAction(ISD::LOAD, MVT::i32, Custom);
180 setOperationAction(ISD::STORE, MVT::i32, Custom);
181
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000182 setTargetDAGCombine(ISD::ADDE);
183 setTargetDAGCombine(ISD::SUBE);
Akira Hatanaka5832fc62013-06-26 18:48:17 +0000184 setTargetDAGCombine(ISD::MUL);
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000185
Daniel Sandersce09d072013-08-28 12:14:50 +0000186 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
Daniel Sanderse6ed5b72013-08-28 12:04:29 +0000187 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
188 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
189
Akira Hatanaka63791212013-09-07 00:52:30 +0000190 if (NoDPLoadStore) {
191 setOperationAction(ISD::LOAD, MVT::f64, Custom);
192 setOperationAction(ISD::STORE, MVT::f64, Custom);
193 }
194
Eric Christopher1c29a652014-07-18 22:55:25 +0000195 if (Subtarget.hasMips32r6()) {
Daniel Sanders308181e2014-06-12 10:44:10 +0000196 // MIPS32r6 replaces the accumulator-based multiplies with a three register
197 // instruction
Daniel Sanders826f8b32014-06-12 10:54:16 +0000198 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
199 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
Daniel Sanders308181e2014-06-12 10:44:10 +0000200 setOperationAction(ISD::MUL, MVT::i32, Legal);
201 setOperationAction(ISD::MULHS, MVT::i32, Legal);
202 setOperationAction(ISD::MULHU, MVT::i32, Legal);
203
204 // MIPS32r6 replaces the accumulator-based division/remainder with separate
205 // three register division and remainder instructions.
206 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
207 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
208 setOperationAction(ISD::SDIV, MVT::i32, Legal);
209 setOperationAction(ISD::UDIV, MVT::i32, Legal);
210 setOperationAction(ISD::SREM, MVT::i32, Legal);
211 setOperationAction(ISD::UREM, MVT::i32, Legal);
Daniel Sanders0fa60412014-06-12 13:39:06 +0000212
213 // MIPS32r6 replaces conditional moves with an equivalent that removes the
214 // need for three GPR read ports.
215 setOperationAction(ISD::SETCC, MVT::i32, Legal);
216 setOperationAction(ISD::SELECT, MVT::i32, Legal);
217 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
218
219 setOperationAction(ISD::SETCC, MVT::f32, Legal);
220 setOperationAction(ISD::SELECT, MVT::f32, Legal);
221 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
222
Eric Christopher1c29a652014-07-18 22:55:25 +0000223 assert(Subtarget.isFP64bit() && "FR=1 is required for MIPS32r6");
Daniel Sanders0fa60412014-06-12 13:39:06 +0000224 setOperationAction(ISD::SETCC, MVT::f64, Legal);
225 setOperationAction(ISD::SELECT, MVT::f64, Legal);
226 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
227
Daniel Sanders3d3ea532014-06-12 15:00:17 +0000228 setOperationAction(ISD::BRCOND, MVT::Other, Legal);
229
Daniel Sanders0fa60412014-06-12 13:39:06 +0000230 // Floating point > and >= are supported via < and <=
231 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
232 setCondCodeAction(ISD::SETOGT, MVT::f32, Expand);
233 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
234 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
235
236 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
237 setCondCodeAction(ISD::SETOGT, MVT::f64, Expand);
238 setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
239 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
Daniel Sanders308181e2014-06-12 10:44:10 +0000240 }
241
Eric Christopher1c29a652014-07-18 22:55:25 +0000242 if (Subtarget.hasMips64r6()) {
Daniel Sanders308181e2014-06-12 10:44:10 +0000243 // MIPS64r6 replaces the accumulator-based multiplies with a three register
244 // instruction
Vasileios Kalintirisef96a8e2015-01-26 12:33:22 +0000245 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
246 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
Daniel Sanders308181e2014-06-12 10:44:10 +0000247 setOperationAction(ISD::MUL, MVT::i64, Legal);
248 setOperationAction(ISD::MULHS, MVT::i64, Legal);
249 setOperationAction(ISD::MULHU, MVT::i64, Legal);
250
251 // MIPS32r6 replaces the accumulator-based division/remainder with separate
252 // three register division and remainder instructions.
253 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
254 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
255 setOperationAction(ISD::SDIV, MVT::i64, Legal);
256 setOperationAction(ISD::UDIV, MVT::i64, Legal);
257 setOperationAction(ISD::SREM, MVT::i64, Legal);
258 setOperationAction(ISD::UREM, MVT::i64, Legal);
Daniel Sanders0fa60412014-06-12 13:39:06 +0000259
260 // MIPS64r6 replaces conditional moves with an equivalent that removes the
261 // need for three GPR read ports.
262 setOperationAction(ISD::SETCC, MVT::i64, Legal);
263 setOperationAction(ISD::SELECT, MVT::i64, Legal);
264 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
Daniel Sanders308181e2014-06-12 10:44:10 +0000265 }
266
Eric Christopher23a3a7c2015-02-26 00:00:24 +0000267 computeRegisterProperties(Subtarget.getRegisterInfo());
Akira Hatanaka96ca1822013-03-13 00:54:29 +0000268}
269
270const MipsTargetLowering *
Eric Christopherb1526602014-09-19 23:30:42 +0000271llvm::createMipsSETargetLowering(const MipsTargetMachine &TM,
Eric Christopher8924d272014-07-18 23:25:04 +0000272 const MipsSubtarget &STI) {
273 return new MipsSETargetLowering(TM, STI);
Akira Hatanaka96ca1822013-03-13 00:54:29 +0000274}
275
Eric Christopherbf33a3c2014-07-02 23:18:40 +0000276const TargetRegisterClass *
277MipsSETargetLowering::getRepRegClassFor(MVT VT) const {
278 if (VT == MVT::Untyped)
Eric Christopher1c29a652014-07-18 22:55:25 +0000279 return Subtarget.hasDSP() ? &Mips::ACC64DSPRegClass : &Mips::ACC64RegClass;
Eric Christopherbf33a3c2014-07-02 23:18:40 +0000280
281 return TargetLowering::getRepRegClassFor(VT);
282}
283
Daniel Sanders7a289d02013-09-23 12:02:46 +0000284// Enable MSA support for the given integer type and Register class.
Daniel Sanders3c9a0ad2013-08-23 10:10:13 +0000285void MipsSETargetLowering::
Daniel Sandersc65f58a2013-09-11 10:15:48 +0000286addMSAIntType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
287 addRegisterClass(Ty, RC);
288
289 // Expand all builtin opcodes.
290 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
291 setOperationAction(Opc, Ty, Expand);
292
293 setOperationAction(ISD::BITCAST, Ty, Legal);
294 setOperationAction(ISD::LOAD, Ty, Legal);
295 setOperationAction(ISD::STORE, Ty, Legal);
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000296 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Custom);
297 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
Daniel Sanders7a289d02013-09-23 12:02:46 +0000298 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
Daniel Sandersc65f58a2013-09-11 10:15:48 +0000299
Daniel Sandersfa5ab1c2013-09-11 10:28:16 +0000300 setOperationAction(ISD::ADD, Ty, Legal);
Daniel Sanders8ca81e42013-09-23 12:57:42 +0000301 setOperationAction(ISD::AND, Ty, Legal);
Daniel Sandersfbcb5822013-09-11 11:58:30 +0000302 setOperationAction(ISD::CTLZ, Ty, Legal);
Daniel Sanders766cb692013-09-23 13:40:21 +0000303 setOperationAction(ISD::CTPOP, Ty, Legal);
Daniel Sandersfbcb5822013-09-11 11:58:30 +0000304 setOperationAction(ISD::MUL, Ty, Legal);
Daniel Sanders8ca81e42013-09-23 12:57:42 +0000305 setOperationAction(ISD::OR, Ty, Legal);
Daniel Sanders607952b2013-09-11 10:38:58 +0000306 setOperationAction(ISD::SDIV, Ty, Legal);
Daniel Sanders0210dd42013-10-01 10:22:35 +0000307 setOperationAction(ISD::SREM, Ty, Legal);
Daniel Sandersfbcb5822013-09-11 11:58:30 +0000308 setOperationAction(ISD::SHL, Ty, Legal);
309 setOperationAction(ISD::SRA, Ty, Legal);
310 setOperationAction(ISD::SRL, Ty, Legal);
311 setOperationAction(ISD::SUB, Ty, Legal);
Daniel Sanders607952b2013-09-11 10:38:58 +0000312 setOperationAction(ISD::UDIV, Ty, Legal);
Daniel Sanders0210dd42013-10-01 10:22:35 +0000313 setOperationAction(ISD::UREM, Ty, Legal);
Daniel Sanderse5087042013-09-24 14:02:15 +0000314 setOperationAction(ISD::VECTOR_SHUFFLE, Ty, Custom);
Daniel Sanderse1d24352013-09-24 12:04:44 +0000315 setOperationAction(ISD::VSELECT, Ty, Legal);
Daniel Sanders8ca81e42013-09-23 12:57:42 +0000316 setOperationAction(ISD::XOR, Ty, Legal);
Daniel Sandersfd538dc2013-09-24 10:46:19 +0000317
Daniel Sanders015972b2013-10-11 10:00:06 +0000318 if (Ty == MVT::v4i32 || Ty == MVT::v2i64) {
319 setOperationAction(ISD::FP_TO_SINT, Ty, Legal);
320 setOperationAction(ISD::FP_TO_UINT, Ty, Legal);
321 setOperationAction(ISD::SINT_TO_FP, Ty, Legal);
322 setOperationAction(ISD::UINT_TO_FP, Ty, Legal);
323 }
324
Daniel Sandersfd538dc2013-09-24 10:46:19 +0000325 setOperationAction(ISD::SETCC, Ty, Legal);
326 setCondCodeAction(ISD::SETNE, Ty, Expand);
327 setCondCodeAction(ISD::SETGE, Ty, Expand);
328 setCondCodeAction(ISD::SETGT, Ty, Expand);
329 setCondCodeAction(ISD::SETUGE, Ty, Expand);
330 setCondCodeAction(ISD::SETUGT, Ty, Expand);
Daniel Sandersc65f58a2013-09-11 10:15:48 +0000331}
332
Daniel Sanders7a289d02013-09-23 12:02:46 +0000333// Enable MSA support for the given floating-point type and Register class.
Daniel Sandersc65f58a2013-09-11 10:15:48 +0000334void MipsSETargetLowering::
335addMSAFloatType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
Daniel Sanders3c9a0ad2013-08-23 10:10:13 +0000336 addRegisterClass(Ty, RC);
Jack Carterbabdcc82013-08-15 12:24:57 +0000337
338 // Expand all builtin opcodes.
339 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
340 setOperationAction(Opc, Ty, Expand);
341
342 setOperationAction(ISD::LOAD, Ty, Legal);
343 setOperationAction(ISD::STORE, Ty, Legal);
344 setOperationAction(ISD::BITCAST, Ty, Legal);
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000345 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal);
Daniel Sandersa5150702013-09-27 12:31:32 +0000346 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
Daniel Sanders1dfddc72013-10-15 13:14:41 +0000347 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
Daniel Sandersf5bd9372013-09-11 10:51:30 +0000348
349 if (Ty != MVT::v8f16) {
Daniel Sanders4f3ff1b2013-09-24 13:02:08 +0000350 setOperationAction(ISD::FABS, Ty, Legal);
Daniel Sandersf5bd9372013-09-11 10:51:30 +0000351 setOperationAction(ISD::FADD, Ty, Legal);
352 setOperationAction(ISD::FDIV, Ty, Legal);
Daniel Sandersa9521602013-10-23 10:36:52 +0000353 setOperationAction(ISD::FEXP2, Ty, Legal);
Daniel Sandersf5bd9372013-09-11 10:51:30 +0000354 setOperationAction(ISD::FLOG2, Ty, Legal);
Daniel Sandersd7103f32013-10-11 10:14:25 +0000355 setOperationAction(ISD::FMA, Ty, Legal);
Daniel Sandersf5bd9372013-09-11 10:51:30 +0000356 setOperationAction(ISD::FMUL, Ty, Legal);
357 setOperationAction(ISD::FRINT, Ty, Legal);
358 setOperationAction(ISD::FSQRT, Ty, Legal);
359 setOperationAction(ISD::FSUB, Ty, Legal);
Daniel Sanderse1d24352013-09-24 12:04:44 +0000360 setOperationAction(ISD::VSELECT, Ty, Legal);
Daniel Sandersfd538dc2013-09-24 10:46:19 +0000361
362 setOperationAction(ISD::SETCC, Ty, Legal);
363 setCondCodeAction(ISD::SETOGE, Ty, Expand);
364 setCondCodeAction(ISD::SETOGT, Ty, Expand);
365 setCondCodeAction(ISD::SETUGE, Ty, Expand);
366 setCondCodeAction(ISD::SETUGT, Ty, Expand);
367 setCondCodeAction(ISD::SETGE, Ty, Expand);
368 setCondCodeAction(ISD::SETGT, Ty, Expand);
Daniel Sandersf5bd9372013-09-11 10:51:30 +0000369 }
Jack Carterbabdcc82013-08-15 12:24:57 +0000370}
Akira Hatanaka96ca1822013-03-13 00:54:29 +0000371
372bool
Matt Arsenault6f2a5262014-07-27 17:46:40 +0000373MipsSETargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
374 unsigned,
375 unsigned,
376 bool *Fast) const {
Akira Hatanaka96ca1822013-03-13 00:54:29 +0000377 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
378
Eric Christopher1c29a652014-07-18 22:55:25 +0000379 if (Subtarget.systemSupportsUnalignedAccess()) {
Daniel Sandersac272632014-05-23 13:18:02 +0000380 // MIPS32r6/MIPS64r6 is required to support unaligned access. It's
381 // implementation defined whether this is handled by hardware, software, or
382 // a hybrid of the two but it's expected that most implementations will
383 // handle the majority of cases in hardware.
384 if (Fast)
385 *Fast = true;
386 return true;
387 }
388
Akira Hatanaka96ca1822013-03-13 00:54:29 +0000389 switch (SVT) {
390 case MVT::i64:
391 case MVT::i32:
392 if (Fast)
393 *Fast = true;
394 return true;
395 default:
396 return false;
397 }
398}
399
Akira Hatanakabe8612f2013-03-30 01:36:35 +0000400SDValue MipsSETargetLowering::LowerOperation(SDValue Op,
401 SelectionDAG &DAG) const {
402 switch(Op.getOpcode()) {
Akira Hatanaka63791212013-09-07 00:52:30 +0000403 case ISD::LOAD: return lowerLOAD(Op, DAG);
404 case ISD::STORE: return lowerSTORE(Op, DAG);
Akira Hatanakabe8612f2013-03-30 01:36:35 +0000405 case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG);
406 case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG);
407 case ISD::MULHS: return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG);
408 case ISD::MULHU: return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG);
409 case ISD::MUL: return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG);
410 case ISD::SDIVREM: return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG);
Akira Hatanakad8fb0322013-04-22 20:13:37 +0000411 case ISD::UDIVREM: return lowerMulDiv(Op, MipsISD::DivRemU, true, true,
412 DAG);
Akira Hatanakaa6bbde52013-04-13 02:13:30 +0000413 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
414 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
Daniel Sanderse6ed5b72013-08-28 12:04:29 +0000415 case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG);
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000416 case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG);
Daniel Sanders7a289d02013-09-23 12:02:46 +0000417 case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG);
Daniel Sanderse5087042013-09-24 14:02:15 +0000418 case ISD::VECTOR_SHUFFLE: return lowerVECTOR_SHUFFLE(Op, DAG);
Akira Hatanakabe8612f2013-03-30 01:36:35 +0000419 }
420
421 return MipsTargetLowering::LowerOperation(Op, DAG);
422}
423
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000424// selectMADD -
425// Transforms a subgraph in CurDAG if the following pattern is found:
426// (addc multLo, Lo0), (adde multHi, Hi0),
427// where,
428// multHi/Lo: product of multiplication
429// Lo0: initial value of Lo register
430// Hi0: initial value of Hi register
431// Return true if pattern matching was successful.
432static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
433 // ADDENode's second operand must be a flag output of an ADDC node in order
434 // for the matching to be successful.
435 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
436
437 if (ADDCNode->getOpcode() != ISD::ADDC)
438 return false;
439
440 SDValue MultHi = ADDENode->getOperand(0);
441 SDValue MultLo = ADDCNode->getOperand(0);
442 SDNode *MultNode = MultHi.getNode();
443 unsigned MultOpc = MultHi.getOpcode();
444
445 // MultHi and MultLo must be generated by the same node,
446 if (MultLo.getNode() != MultNode)
447 return false;
448
449 // and it must be a multiplication.
450 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
451 return false;
452
453 // MultLo amd MultHi must be the first and second output of MultNode
454 // respectively.
455 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
456 return false;
457
458 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
459 // of the values of MultNode, in which case MultNode will be removed in later
460 // phases.
461 // If there exist users other than ADDENode or ADDCNode, this function returns
462 // here, which will result in MultNode being mapped to a single MULT
463 // instruction node rather than a pair of MULT and MADD instructions being
464 // produced.
465 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
466 return false;
467
Andrew Trickef9de2a2013-05-25 02:42:55 +0000468 SDLoc DL(ADDENode);
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000469
470 // Initialize accumulator.
Akira Hatanakad98c99f2013-10-15 01:12:50 +0000471 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000472 ADDCNode->getOperand(1),
473 ADDENode->getOperand(1));
474
475 // create MipsMAdd(u) node
476 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
477
478 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
479 MultNode->getOperand(0),// Factor 0
480 MultNode->getOperand(1),// Factor 1
481 ACCIn);
482
483 // replace uses of adde and addc here
484 if (!SDValue(ADDCNode, 0).use_empty()) {
Akira Hatanakad98c99f2013-10-15 01:12:50 +0000485 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MAdd);
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000486 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
487 }
488 if (!SDValue(ADDENode, 0).use_empty()) {
Akira Hatanakad98c99f2013-10-15 01:12:50 +0000489 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MAdd);
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000490 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
491 }
492
493 return true;
494}
495
496// selectMSUB -
497// Transforms a subgraph in CurDAG if the following pattern is found:
498// (addc Lo0, multLo), (sube Hi0, multHi),
499// where,
500// multHi/Lo: product of multiplication
501// Lo0: initial value of Lo register
502// Hi0: initial value of Hi register
503// Return true if pattern matching was successful.
504static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
505 // SUBENode's second operand must be a flag output of an SUBC node in order
506 // for the matching to be successful.
507 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
508
509 if (SUBCNode->getOpcode() != ISD::SUBC)
510 return false;
511
512 SDValue MultHi = SUBENode->getOperand(1);
513 SDValue MultLo = SUBCNode->getOperand(1);
514 SDNode *MultNode = MultHi.getNode();
515 unsigned MultOpc = MultHi.getOpcode();
516
517 // MultHi and MultLo must be generated by the same node,
518 if (MultLo.getNode() != MultNode)
519 return false;
520
521 // and it must be a multiplication.
522 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
523 return false;
524
525 // MultLo amd MultHi must be the first and second output of MultNode
526 // respectively.
527 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
528 return false;
529
530 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
531 // of the values of MultNode, in which case MultNode will be removed in later
532 // phases.
533 // If there exist users other than SUBENode or SUBCNode, this function returns
534 // here, which will result in MultNode being mapped to a single MULT
535 // instruction node rather than a pair of MULT and MSUB instructions being
536 // produced.
537 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
538 return false;
539
Andrew Trickef9de2a2013-05-25 02:42:55 +0000540 SDLoc DL(SUBENode);
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000541
542 // Initialize accumulator.
Akira Hatanakad98c99f2013-10-15 01:12:50 +0000543 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000544 SUBCNode->getOperand(0),
545 SUBENode->getOperand(0));
546
547 // create MipsSub(u) node
548 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
549
550 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
551 MultNode->getOperand(0),// Factor 0
552 MultNode->getOperand(1),// Factor 1
553 ACCIn);
554
555 // replace uses of sube and subc here
556 if (!SDValue(SUBCNode, 0).use_empty()) {
Akira Hatanakad98c99f2013-10-15 01:12:50 +0000557 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MSub);
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000558 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
559 }
560 if (!SDValue(SUBENode, 0).use_empty()) {
Akira Hatanakad98c99f2013-10-15 01:12:50 +0000561 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MSub);
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000562 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
563 }
564
565 return true;
566}
567
568static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
569 TargetLowering::DAGCombinerInfo &DCI,
Eric Christopher1c29a652014-07-18 22:55:25 +0000570 const MipsSubtarget &Subtarget) {
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000571 if (DCI.isBeforeLegalize())
572 return SDValue();
573
Eric Christopher1c29a652014-07-18 22:55:25 +0000574 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() &&
Daniel Sanders826f8b32014-06-12 10:54:16 +0000575 N->getValueType(0) == MVT::i32 && selectMADD(N, &DAG))
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000576 return SDValue(N, 0);
577
578 return SDValue();
579}
580
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000581// Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT
582//
583// Performs the following transformations:
584// - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its
585// sign/zero-extension is completely overwritten by the new one performed by
586// the ISD::AND.
587// - Removes redundant zero extensions performed by an ISD::AND.
588static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
589 TargetLowering::DAGCombinerInfo &DCI,
Eric Christopher1c29a652014-07-18 22:55:25 +0000590 const MipsSubtarget &Subtarget) {
591 if (!Subtarget.hasMSA())
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000592 return SDValue();
593
594 SDValue Op0 = N->getOperand(0);
595 SDValue Op1 = N->getOperand(1);
596 unsigned Op0Opcode = Op0->getOpcode();
597
598 // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d)
599 // where $d + 1 == 2^n and n == 32
600 // or $d + 1 == 2^n and n <= 32 and ZExt
601 // -> (MipsVExtractZExt $a, $b, $c)
602 if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT ||
603 Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) {
604 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1);
605
606 if (!Mask)
607 return SDValue();
608
609 int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2();
610
611 if (Log2IfPositive <= 0)
612 return SDValue(); // Mask+1 is not a power of 2
613
614 SDValue Op0Op2 = Op0->getOperand(2);
615 EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT();
616 unsigned ExtendTySize = ExtendTy.getSizeInBits();
617 unsigned Log2 = Log2IfPositive;
618
619 if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) ||
620 Log2 == ExtendTySize) {
621 SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
Chandler Carruth356665a2014-08-01 22:09:43 +0000622 return DAG.getNode(MipsISD::VEXTRACT_ZEXT_ELT, SDLoc(Op0),
623 Op0->getVTList(),
624 makeArrayRef(Ops, Op0->getNumOperands()));
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000625 }
626 }
627
628 return SDValue();
629}
630
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000631// Determine if the specified node is a constant vector splat.
632//
633// Returns true and sets Imm if:
634// * N is a ISD::BUILD_VECTOR representing a constant splat
635//
636// This function is quite similar to MipsSEDAGToDAGISel::selectVSplat. The
637// differences are that it assumes the MSA has already been checked and the
638// arbitrary requirement for a maximum of 32-bit integers isn't applied (and
639// must not be in order for binsri.d to be selectable).
640static bool isVSplat(SDValue N, APInt &Imm, bool IsLittleEndian) {
641 BuildVectorSDNode *Node = dyn_cast<BuildVectorSDNode>(N.getNode());
642
Craig Topper062a2ba2014-04-25 05:30:21 +0000643 if (!Node)
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000644 return false;
645
646 APInt SplatValue, SplatUndef;
647 unsigned SplatBitSize;
648 bool HasAnyUndefs;
649
650 if (!Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
651 8, !IsLittleEndian))
652 return false;
653
654 Imm = SplatValue;
655
656 return true;
657}
658
Daniel Sandersab94b532013-10-30 15:20:38 +0000659// Test whether the given node is an all-ones build_vector.
660static bool isVectorAllOnes(SDValue N) {
661 // Look through bitcasts. Endianness doesn't matter because we are looking
662 // for an all-ones value.
663 if (N->getOpcode() == ISD::BITCAST)
664 N = N->getOperand(0);
665
666 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
667
668 if (!BVN)
669 return false;
670
671 APInt SplatValue, SplatUndef;
672 unsigned SplatBitSize;
673 bool HasAnyUndefs;
674
675 // Endianness doesn't matter in this context because we are looking for
676 // an all-ones value.
677 if (BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs))
678 return SplatValue.isAllOnesValue();
679
680 return false;
681}
682
683// Test whether N is the bitwise inverse of OfNode.
684static bool isBitwiseInverse(SDValue N, SDValue OfNode) {
685 if (N->getOpcode() != ISD::XOR)
686 return false;
687
688 if (isVectorAllOnes(N->getOperand(0)))
689 return N->getOperand(1) == OfNode;
690
691 if (isVectorAllOnes(N->getOperand(1)))
692 return N->getOperand(0) == OfNode;
693
694 return false;
695}
696
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000697// Perform combines where ISD::OR is the root node.
698//
699// Performs the following transformations:
700// - (or (and $a, $mask), (and $b, $inv_mask)) => (vselect $mask, $a, $b)
701// where $inv_mask is the bitwise inverse of $mask and the 'or' has a 128-bit
702// vector type.
703static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
704 TargetLowering::DAGCombinerInfo &DCI,
Eric Christopher1c29a652014-07-18 22:55:25 +0000705 const MipsSubtarget &Subtarget) {
706 if (!Subtarget.hasMSA())
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000707 return SDValue();
708
709 EVT Ty = N->getValueType(0);
710
711 if (!Ty.is128BitVector())
712 return SDValue();
713
714 SDValue Op0 = N->getOperand(0);
715 SDValue Op1 = N->getOperand(1);
716
717 if (Op0->getOpcode() == ISD::AND && Op1->getOpcode() == ISD::AND) {
718 SDValue Op0Op0 = Op0->getOperand(0);
719 SDValue Op0Op1 = Op0->getOperand(1);
720 SDValue Op1Op0 = Op1->getOperand(0);
721 SDValue Op1Op1 = Op1->getOperand(1);
Eric Christopher1c29a652014-07-18 22:55:25 +0000722 bool IsLittleEndian = !Subtarget.isLittle();
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000723
724 SDValue IfSet, IfClr, Cond;
Daniel Sandersab94b532013-10-30 15:20:38 +0000725 bool IsConstantMask = false;
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000726 APInt Mask, InvMask;
727
728 // If Op0Op0 is an appropriate mask, try to find it's inverse in either
729 // Op1Op0, or Op1Op1. Keep track of the Cond, IfSet, and IfClr nodes, while
730 // looking.
731 // IfClr will be set if we find a valid match.
732 if (isVSplat(Op0Op0, Mask, IsLittleEndian)) {
733 Cond = Op0Op0;
734 IfSet = Op0Op1;
735
Daniel Sandersc8c50fb2013-11-21 16:11:31 +0000736 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
737 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000738 IfClr = Op1Op1;
Daniel Sandersc8c50fb2013-11-21 16:11:31 +0000739 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
740 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000741 IfClr = Op1Op0;
Daniel Sandersab94b532013-10-30 15:20:38 +0000742
743 IsConstantMask = true;
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000744 }
745
746 // If IfClr is not yet set, and Op0Op1 is an appropriate mask, try the same
747 // thing again using this mask.
748 // IfClr will be set if we find a valid match.
749 if (!IfClr.getNode() && isVSplat(Op0Op1, Mask, IsLittleEndian)) {
750 Cond = Op0Op1;
751 IfSet = Op0Op0;
752
Daniel Sandersc8c50fb2013-11-21 16:11:31 +0000753 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
754 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000755 IfClr = Op1Op1;
Daniel Sandersc8c50fb2013-11-21 16:11:31 +0000756 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
757 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000758 IfClr = Op1Op0;
Daniel Sandersab94b532013-10-30 15:20:38 +0000759
760 IsConstantMask = true;
761 }
762
763 // If IfClr is not yet set, try looking for a non-constant match.
764 // IfClr will be set if we find a valid match amongst the eight
765 // possibilities.
766 if (!IfClr.getNode()) {
767 if (isBitwiseInverse(Op0Op0, Op1Op0)) {
768 Cond = Op1Op0;
769 IfSet = Op1Op1;
770 IfClr = Op0Op1;
771 } else if (isBitwiseInverse(Op0Op1, Op1Op0)) {
772 Cond = Op1Op0;
773 IfSet = Op1Op1;
774 IfClr = Op0Op0;
775 } else if (isBitwiseInverse(Op0Op0, Op1Op1)) {
776 Cond = Op1Op1;
777 IfSet = Op1Op0;
778 IfClr = Op0Op1;
779 } else if (isBitwiseInverse(Op0Op1, Op1Op1)) {
780 Cond = Op1Op1;
781 IfSet = Op1Op0;
782 IfClr = Op0Op0;
783 } else if (isBitwiseInverse(Op1Op0, Op0Op0)) {
784 Cond = Op0Op0;
785 IfSet = Op0Op1;
786 IfClr = Op1Op1;
787 } else if (isBitwiseInverse(Op1Op1, Op0Op0)) {
788 Cond = Op0Op0;
789 IfSet = Op0Op1;
790 IfClr = Op1Op0;
791 } else if (isBitwiseInverse(Op1Op0, Op0Op1)) {
792 Cond = Op0Op1;
793 IfSet = Op0Op0;
794 IfClr = Op1Op1;
795 } else if (isBitwiseInverse(Op1Op1, Op0Op1)) {
796 Cond = Op0Op1;
797 IfSet = Op0Op0;
798 IfClr = Op1Op0;
799 }
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000800 }
801
802 // At this point, IfClr will be set if we have a valid match.
803 if (!IfClr.getNode())
804 return SDValue();
805
806 assert(Cond.getNode() && IfSet.getNode());
807
808 // Fold degenerate cases.
Daniel Sandersab94b532013-10-30 15:20:38 +0000809 if (IsConstantMask) {
810 if (Mask.isAllOnesValue())
811 return IfSet;
812 else if (Mask == 0)
813 return IfClr;
814 }
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000815
816 // Transform the DAG into an equivalent VSELECT.
Daniel Sandersdf2215452014-03-12 11:54:00 +0000817 return DAG.getNode(ISD::VSELECT, SDLoc(N), Ty, Cond, IfSet, IfClr);
Daniel Sanders53fe6c42013-10-30 13:51:01 +0000818 }
819
820 return SDValue();
821}
822
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000823static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
824 TargetLowering::DAGCombinerInfo &DCI,
Eric Christopher1c29a652014-07-18 22:55:25 +0000825 const MipsSubtarget &Subtarget) {
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000826 if (DCI.isBeforeLegalize())
827 return SDValue();
828
Eric Christopher1c29a652014-07-18 22:55:25 +0000829 if (Subtarget.hasMips32() && N->getValueType(0) == MVT::i32 &&
Akira Hatanaka9efcd762013-03-30 01:42:24 +0000830 selectMSUB(N, &DAG))
831 return SDValue(N, 0);
832
833 return SDValue();
834}
835
Benjamin Kramerbdc49562016-06-12 15:39:02 +0000836static SDValue genConstMult(SDValue X, uint64_t C, const SDLoc &DL, EVT VT,
Akira Hatanaka5832fc62013-06-26 18:48:17 +0000837 EVT ShiftTy, SelectionDAG &DAG) {
838 // Clear the upper (64 - VT.sizeInBits) bits.
839 C &= ((uint64_t)-1) >> (64 - VT.getSizeInBits());
840
841 // Return 0.
842 if (C == 0)
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +0000843 return DAG.getConstant(0, DL, VT);
Akira Hatanaka5832fc62013-06-26 18:48:17 +0000844
845 // Return x.
846 if (C == 1)
847 return X;
848
849 // If c is power of 2, return (shl x, log2(c)).
850 if (isPowerOf2_64(C))
851 return DAG.getNode(ISD::SHL, DL, VT, X,
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +0000852 DAG.getConstant(Log2_64(C), DL, ShiftTy));
Akira Hatanaka5832fc62013-06-26 18:48:17 +0000853
854 unsigned Log2Ceil = Log2_64_Ceil(C);
855 uint64_t Floor = 1LL << Log2_64(C);
856 uint64_t Ceil = Log2Ceil == 64 ? 0LL : 1LL << Log2Ceil;
857
858 // If |c - floor_c| <= |c - ceil_c|,
859 // where floor_c = pow(2, floor(log2(c))) and ceil_c = pow(2, ceil(log2(c))),
860 // return (add constMult(x, floor_c), constMult(x, c - floor_c)).
861 if (C - Floor <= Ceil - C) {
862 SDValue Op0 = genConstMult(X, Floor, DL, VT, ShiftTy, DAG);
863 SDValue Op1 = genConstMult(X, C - Floor, DL, VT, ShiftTy, DAG);
864 return DAG.getNode(ISD::ADD, DL, VT, Op0, Op1);
865 }
866
867 // If |c - floor_c| > |c - ceil_c|,
868 // return (sub constMult(x, ceil_c), constMult(x, ceil_c - c)).
869 SDValue Op0 = genConstMult(X, Ceil, DL, VT, ShiftTy, DAG);
870 SDValue Op1 = genConstMult(X, Ceil - C, DL, VT, ShiftTy, DAG);
871 return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1);
872}
873
874static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
875 const TargetLowering::DAGCombinerInfo &DCI,
876 const MipsSETargetLowering *TL) {
877 EVT VT = N->getValueType(0);
878
879 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
880 if (!VT.isVector())
Mehdi Amini9639d652015-07-09 02:09:20 +0000881 return genConstMult(N->getOperand(0), C->getZExtValue(), SDLoc(N), VT,
Mehdi Aminieaabc512015-07-09 15:12:23 +0000882 TL->getScalarShiftAmountTy(DAG.getDataLayout(), VT),
883 DAG);
Akira Hatanaka5832fc62013-06-26 18:48:17 +0000884
885 return SDValue(N, 0);
886}
887
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000888static SDValue performDSPShiftCombine(unsigned Opc, SDNode *N, EVT Ty,
889 SelectionDAG &DAG,
Eric Christopher1c29a652014-07-18 22:55:25 +0000890 const MipsSubtarget &Subtarget) {
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000891 // See if this is a vector splat immediate node.
892 APInt SplatValue, SplatUndef;
893 unsigned SplatBitSize;
894 bool HasAnyUndefs;
Sanjay Patel1ed771f2016-09-14 16:37:15 +0000895 unsigned EltSize = Ty.getScalarSizeInBits();
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000896 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
897
Eric Christopher1c29a652014-07-18 22:55:25 +0000898 if (!Subtarget.hasDSP())
Daniel Sanders6e664bc2013-11-21 11:40:14 +0000899 return SDValue();
900
Akira Hatanaka0d6964c2013-04-22 19:58:23 +0000901 if (!BV ||
Akira Hatanakad8fb0322013-04-22 20:13:37 +0000902 !BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
Eric Christopher1c29a652014-07-18 22:55:25 +0000903 EltSize, !Subtarget.isLittle()) ||
Akira Hatanaka0d6964c2013-04-22 19:58:23 +0000904 (SplatBitSize != EltSize) ||
Akira Hatanakae9d0b312013-04-23 18:09:42 +0000905 (SplatValue.getZExtValue() >= EltSize))
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000906 return SDValue();
907
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +0000908 SDLoc DL(N);
909 return DAG.getNode(Opc, DL, Ty, N->getOperand(0),
910 DAG.getConstant(SplatValue.getZExtValue(), DL, MVT::i32));
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000911}
912
913static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
914 TargetLowering::DAGCombinerInfo &DCI,
Eric Christopher1c29a652014-07-18 22:55:25 +0000915 const MipsSubtarget &Subtarget) {
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000916 EVT Ty = N->getValueType(0);
917
918 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
919 return SDValue();
920
921 return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget);
922}
923
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000924// Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold
925// constant splats into MipsISD::SHRA_DSP for DSPr2.
926//
927// Performs the following transformations:
928// - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its
929// sign/zero-extension is completely overwritten by the new one performed by
930// the ISD::SRA and ISD::SHL nodes.
931// - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL
932// sequence.
933//
934// See performDSPShiftCombine for more information about the transformation
935// used for DSPr2.
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000936static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
937 TargetLowering::DAGCombinerInfo &DCI,
Eric Christopher1c29a652014-07-18 22:55:25 +0000938 const MipsSubtarget &Subtarget) {
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000939 EVT Ty = N->getValueType(0);
940
Eric Christopher1c29a652014-07-18 22:55:25 +0000941 if (Subtarget.hasMSA()) {
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000942 SDValue Op0 = N->getOperand(0);
943 SDValue Op1 = N->getOperand(1);
944
945 // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d)
946 // where $d + sizeof($c) == 32
947 // or $d + sizeof($c) <= 32 and SExt
948 // -> (MipsVExtractSExt $a, $b, $c)
949 if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) {
950 SDValue Op0Op0 = Op0->getOperand(0);
951 ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1);
952
953 if (!ShAmount)
954 return SDValue();
955
Daniel Sandersf4f1a872013-09-27 09:25:29 +0000956 if (Op0Op0->getOpcode() != MipsISD::VEXTRACT_SEXT_ELT &&
957 Op0Op0->getOpcode() != MipsISD::VEXTRACT_ZEXT_ELT)
958 return SDValue();
959
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000960 EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT();
961 unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits();
962
963 if (TotalBits == 32 ||
964 (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT &&
965 TotalBits <= 32)) {
966 SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
967 Op0Op0->getOperand(2) };
Chandler Carruth356665a2014-08-01 22:09:43 +0000968 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, SDLoc(Op0Op0),
969 Op0Op0->getVTList(),
970 makeArrayRef(Ops, Op0Op0->getNumOperands()));
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +0000971 }
972 }
973 }
974
Eric Christopher1c29a652014-07-18 22:55:25 +0000975 if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget.hasDSPR2()))
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000976 return SDValue();
977
978 return performDSPShiftCombine(MipsISD::SHRA_DSP, N, Ty, DAG, Subtarget);
979}
980
981
982static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
983 TargetLowering::DAGCombinerInfo &DCI,
Eric Christopher1c29a652014-07-18 22:55:25 +0000984 const MipsSubtarget &Subtarget) {
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000985 EVT Ty = N->getValueType(0);
986
Eric Christopher1c29a652014-07-18 22:55:25 +0000987 if (((Ty != MVT::v2i16) || !Subtarget.hasDSPR2()) && (Ty != MVT::v4i8))
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +0000988 return SDValue();
989
990 return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget);
991}
992
Akira Hatanaka68741cc2013-04-30 22:37:26 +0000993static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) {
994 bool IsV216 = (Ty == MVT::v2i16);
995
996 switch (CC) {
997 case ISD::SETEQ:
998 case ISD::SETNE: return true;
999 case ISD::SETLT:
1000 case ISD::SETLE:
1001 case ISD::SETGT:
1002 case ISD::SETGE: return IsV216;
1003 case ISD::SETULT:
1004 case ISD::SETULE:
1005 case ISD::SETUGT:
1006 case ISD::SETUGE: return !IsV216;
1007 default: return false;
1008 }
1009}
1010
1011static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) {
1012 EVT Ty = N->getValueType(0);
1013
1014 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
1015 return SDValue();
1016
1017 if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get()))
1018 return SDValue();
1019
Andrew Trickef9de2a2013-05-25 02:42:55 +00001020 return DAG.getNode(MipsISD::SETCC_DSP, SDLoc(N), Ty, N->getOperand(0),
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001021 N->getOperand(1), N->getOperand(2));
1022}
1023
1024static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
1025 EVT Ty = N->getValueType(0);
1026
Daniel Sanders3ce56622013-09-24 12:18:31 +00001027 if (Ty.is128BitVector() && Ty.isInteger()) {
1028 // Try the following combines:
1029 // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b)
1030 // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b)
1031 // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b)
1032 // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b)
1033 // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b)
1034 // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b)
1035 // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b)
1036 // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b)
1037 // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but
1038 // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the
1039 // legalizer.
1040 SDValue Op0 = N->getOperand(0);
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001041
Daniel Sanders3ce56622013-09-24 12:18:31 +00001042 if (Op0->getOpcode() != ISD::SETCC)
1043 return SDValue();
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001044
Daniel Sanders3ce56622013-09-24 12:18:31 +00001045 ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get();
1046 bool Signed;
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001047
Daniel Sanders3ce56622013-09-24 12:18:31 +00001048 if (CondCode == ISD::SETLT || CondCode == ISD::SETLE)
1049 Signed = true;
1050 else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE)
1051 Signed = false;
1052 else
1053 return SDValue();
1054
1055 SDValue Op1 = N->getOperand(1);
1056 SDValue Op2 = N->getOperand(2);
1057 SDValue Op0Op0 = Op0->getOperand(0);
1058 SDValue Op0Op1 = Op0->getOperand(1);
1059
1060 if (Op1 == Op0Op0 && Op2 == Op0Op1)
1061 return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N),
1062 Ty, Op1, Op2);
1063 else if (Op1 == Op0Op1 && Op2 == Op0Op0)
1064 return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N),
1065 Ty, Op1, Op2);
1066 } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) {
1067 SDValue SetCC = N->getOperand(0);
1068
1069 if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
1070 return SDValue();
1071
1072 return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
1073 SetCC.getOperand(0), SetCC.getOperand(1),
1074 N->getOperand(1), N->getOperand(2), SetCC.getOperand(2));
1075 }
1076
1077 return SDValue();
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001078}
1079
Daniel Sandersf7456c72013-09-23 13:22:24 +00001080static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
Eric Christopher1c29a652014-07-18 22:55:25 +00001081 const MipsSubtarget &Subtarget) {
Daniel Sandersf7456c72013-09-23 13:22:24 +00001082 EVT Ty = N->getValueType(0);
1083
Eric Christopher1c29a652014-07-18 22:55:25 +00001084 if (Subtarget.hasMSA() && Ty.is128BitVector() && Ty.isInteger()) {
Daniel Sandersf7456c72013-09-23 13:22:24 +00001085 // Try the following combines:
1086 // (xor (or $a, $b), (build_vector allones))
1087 // (xor (or $a, $b), (bitcast (build_vector allones)))
1088 SDValue Op0 = N->getOperand(0);
1089 SDValue Op1 = N->getOperand(1);
1090 SDValue NotOp;
Daniel Sandersf7456c72013-09-23 13:22:24 +00001091
1092 if (ISD::isBuildVectorAllOnes(Op0.getNode()))
1093 NotOp = Op1;
1094 else if (ISD::isBuildVectorAllOnes(Op1.getNode()))
1095 NotOp = Op0;
Daniel Sandersf7456c72013-09-23 13:22:24 +00001096 else
1097 return SDValue();
1098
1099 if (NotOp->getOpcode() == ISD::OR)
1100 return DAG.getNode(MipsISD::VNOR, SDLoc(N), Ty, NotOp->getOperand(0),
1101 NotOp->getOperand(1));
1102 }
1103
1104 return SDValue();
1105}
1106
Akira Hatanaka9efcd762013-03-30 01:42:24 +00001107SDValue
1108MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1109 SelectionDAG &DAG = DCI.DAG;
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001110 SDValue Val;
Akira Hatanaka9efcd762013-03-30 01:42:24 +00001111
1112 switch (N->getOpcode()) {
1113 case ISD::ADDE:
1114 return performADDECombine(N, DAG, DCI, Subtarget);
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +00001115 case ISD::AND:
1116 Val = performANDCombine(N, DAG, DCI, Subtarget);
1117 break;
Daniel Sanders53fe6c42013-10-30 13:51:01 +00001118 case ISD::OR:
1119 Val = performORCombine(N, DAG, DCI, Subtarget);
1120 break;
Akira Hatanaka9efcd762013-03-30 01:42:24 +00001121 case ISD::SUBE:
1122 return performSUBECombine(N, DAG, DCI, Subtarget);
Akira Hatanaka5832fc62013-06-26 18:48:17 +00001123 case ISD::MUL:
1124 return performMULCombine(N, DAG, DCI, this);
Akira Hatanaka1ebb2a12013-04-19 23:21:32 +00001125 case ISD::SHL:
1126 return performSHLCombine(N, DAG, DCI, Subtarget);
1127 case ISD::SRA:
1128 return performSRACombine(N, DAG, DCI, Subtarget);
1129 case ISD::SRL:
1130 return performSRLCombine(N, DAG, DCI, Subtarget);
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001131 case ISD::VSELECT:
1132 return performVSELECTCombine(N, DAG);
Daniel Sandersf7456c72013-09-23 13:22:24 +00001133 case ISD::XOR:
1134 Val = performXORCombine(N, DAG, Subtarget);
1135 break;
1136 case ISD::SETCC:
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001137 Val = performSETCCCombine(N, DAG);
1138 break;
Akira Hatanaka9efcd762013-03-30 01:42:24 +00001139 }
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001140
Daniel Sanders62aeab82013-10-30 13:31:27 +00001141 if (Val.getNode()) {
1142 DEBUG(dbgs() << "\nMipsSE DAG Combine:\n";
1143 N->printrWithDepth(dbgs(), &DAG);
1144 dbgs() << "\n=> \n";
1145 Val.getNode()->printrWithDepth(dbgs(), &DAG);
1146 dbgs() << "\n");
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001147 return Val;
Daniel Sanders62aeab82013-10-30 13:31:27 +00001148 }
Akira Hatanaka68741cc2013-04-30 22:37:26 +00001149
1150 return MipsTargetLowering::PerformDAGCombine(N, DCI);
Akira Hatanaka9efcd762013-03-30 01:42:24 +00001151}
1152
Akira Hatanaka96ca1822013-03-13 00:54:29 +00001153MachineBasicBlock *
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00001154MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
Akira Hatanaka96ca1822013-03-13 00:54:29 +00001155 MachineBasicBlock *BB) const {
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00001156 switch (MI.getOpcode()) {
Akira Hatanaka96ca1822013-03-13 00:54:29 +00001157 default:
1158 return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB);
1159 case Mips::BPOSGE32_PSEUDO:
1160 return emitBPOSGE32(MI, BB);
Daniel Sandersce09d072013-08-28 12:14:50 +00001161 case Mips::SNZ_B_PSEUDO:
1162 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_B);
1163 case Mips::SNZ_H_PSEUDO:
1164 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_H);
1165 case Mips::SNZ_W_PSEUDO:
1166 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_W);
1167 case Mips::SNZ_D_PSEUDO:
1168 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_D);
1169 case Mips::SNZ_V_PSEUDO:
1170 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_V);
1171 case Mips::SZ_B_PSEUDO:
1172 return emitMSACBranchPseudo(MI, BB, Mips::BZ_B);
1173 case Mips::SZ_H_PSEUDO:
1174 return emitMSACBranchPseudo(MI, BB, Mips::BZ_H);
1175 case Mips::SZ_W_PSEUDO:
1176 return emitMSACBranchPseudo(MI, BB, Mips::BZ_W);
1177 case Mips::SZ_D_PSEUDO:
1178 return emitMSACBranchPseudo(MI, BB, Mips::BZ_D);
1179 case Mips::SZ_V_PSEUDO:
1180 return emitMSACBranchPseudo(MI, BB, Mips::BZ_V);
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00001181 case Mips::COPY_FW_PSEUDO:
1182 return emitCOPY_FW(MI, BB);
1183 case Mips::COPY_FD_PSEUDO:
1184 return emitCOPY_FD(MI, BB);
Daniel Sandersa5150702013-09-27 12:31:32 +00001185 case Mips::INSERT_FW_PSEUDO:
1186 return emitINSERT_FW(MI, BB);
1187 case Mips::INSERT_FD_PSEUDO:
1188 return emitINSERT_FD(MI, BB);
Daniel Sanderse296a0f2014-04-30 12:09:32 +00001189 case Mips::INSERT_B_VIDX_PSEUDO:
Daniel Sanderseda60d22015-05-05 10:32:24 +00001190 case Mips::INSERT_B_VIDX64_PSEUDO:
Daniel Sanderse296a0f2014-04-30 12:09:32 +00001191 return emitINSERT_DF_VIDX(MI, BB, 1, false);
1192 case Mips::INSERT_H_VIDX_PSEUDO:
Daniel Sanderseda60d22015-05-05 10:32:24 +00001193 case Mips::INSERT_H_VIDX64_PSEUDO:
Daniel Sanderse296a0f2014-04-30 12:09:32 +00001194 return emitINSERT_DF_VIDX(MI, BB, 2, false);
1195 case Mips::INSERT_W_VIDX_PSEUDO:
Daniel Sanderseda60d22015-05-05 10:32:24 +00001196 case Mips::INSERT_W_VIDX64_PSEUDO:
Daniel Sanderse296a0f2014-04-30 12:09:32 +00001197 return emitINSERT_DF_VIDX(MI, BB, 4, false);
1198 case Mips::INSERT_D_VIDX_PSEUDO:
Daniel Sanderseda60d22015-05-05 10:32:24 +00001199 case Mips::INSERT_D_VIDX64_PSEUDO:
Daniel Sanderse296a0f2014-04-30 12:09:32 +00001200 return emitINSERT_DF_VIDX(MI, BB, 8, false);
1201 case Mips::INSERT_FW_VIDX_PSEUDO:
Daniel Sanderseda60d22015-05-05 10:32:24 +00001202 case Mips::INSERT_FW_VIDX64_PSEUDO:
Daniel Sanderse296a0f2014-04-30 12:09:32 +00001203 return emitINSERT_DF_VIDX(MI, BB, 4, true);
1204 case Mips::INSERT_FD_VIDX_PSEUDO:
Daniel Sanderseda60d22015-05-05 10:32:24 +00001205 case Mips::INSERT_FD_VIDX64_PSEUDO:
Daniel Sanderse296a0f2014-04-30 12:09:32 +00001206 return emitINSERT_DF_VIDX(MI, BB, 8, true);
Daniel Sanders1dfddc72013-10-15 13:14:41 +00001207 case Mips::FILL_FW_PSEUDO:
1208 return emitFILL_FW(MI, BB);
1209 case Mips::FILL_FD_PSEUDO:
1210 return emitFILL_FD(MI, BB);
Daniel Sandersa9521602013-10-23 10:36:52 +00001211 case Mips::FEXP2_W_1_PSEUDO:
1212 return emitFEXP2_W_1(MI, BB);
1213 case Mips::FEXP2_D_1_PSEUDO:
1214 return emitFEXP2_D_1(MI, BB);
Simon Dardis0e2ee3b2016-11-18 16:17:44 +00001215 case Mips::ST_F16:
1216 return emitST_F16_PSEUDO(MI, BB);
1217 case Mips::LD_F16:
1218 return emitLD_F16_PSEUDO(MI, BB);
1219 case Mips::MSA_FP_EXTEND_W_PSEUDO:
1220 return emitFPEXTEND_PSEUDO(MI, BB, false);
1221 case Mips::MSA_FP_ROUND_W_PSEUDO:
1222 return emitFPROUND_PSEUDO(MI, BB, false);
1223 case Mips::MSA_FP_EXTEND_D_PSEUDO:
1224 return emitFPEXTEND_PSEUDO(MI, BB, true);
1225 case Mips::MSA_FP_ROUND_D_PSEUDO:
1226 return emitFPROUND_PSEUDO(MI, BB, true);
Akira Hatanaka96ca1822013-03-13 00:54:29 +00001227 }
1228}
1229
Daniel Sanders23e98772014-11-02 16:09:29 +00001230bool MipsSETargetLowering::isEligibleForTailCallOptimization(
1231 const CCState &CCInfo, unsigned NextStackOffset,
1232 const MipsFunctionInfo &FI) const {
Simon Dardis57f4ae42016-08-04 09:17:07 +00001233 if (!UseMipsTailCalls)
Akira Hatanaka96ca1822013-03-13 00:54:29 +00001234 return false;
1235
Vasileios Kalintiris43dff0c2015-10-26 12:38:43 +00001236 // Exception has to be cleared with eret.
1237 if (FI.isISR())
1238 return false;
1239
Akira Hatanaka96ca1822013-03-13 00:54:29 +00001240 // Return false if either the callee or caller has a byval argument.
Daniel Sanders23e98772014-11-02 16:09:29 +00001241 if (CCInfo.getInRegsParamsCount() > 0 || FI.hasByvalArg())
Akira Hatanaka96ca1822013-03-13 00:54:29 +00001242 return false;
1243
1244 // Return true if the callee's argument area is no larger than the
1245 // caller's.
1246 return NextStackOffset <= FI.getIncomingArgSize();
1247}
1248
1249void MipsSETargetLowering::
1250getOpndList(SmallVectorImpl<SDValue> &Ops,
1251 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
1252 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
Sasa Stankovic7072a792014-10-01 08:22:21 +00001253 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee,
1254 SDValue Chain) const {
Akira Hatanaka168d4e52013-11-27 23:38:42 +00001255 Ops.push_back(Callee);
Akira Hatanaka96ca1822013-03-13 00:54:29 +00001256 MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal,
Sasa Stankovic7072a792014-10-01 08:22:21 +00001257 InternalLinkage, IsCallReloc, CLI, Callee,
1258 Chain);
Akira Hatanaka96ca1822013-03-13 00:54:29 +00001259}
1260
Akira Hatanaka63791212013-09-07 00:52:30 +00001261SDValue MipsSETargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1262 LoadSDNode &Nd = *cast<LoadSDNode>(Op);
1263
1264 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1265 return MipsTargetLowering::lowerLOAD(Op, DAG);
1266
1267 // Replace a double precision load with two i32 loads and a buildpair64.
1268 SDLoc DL(Op);
1269 SDValue Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1270 EVT PtrVT = Ptr.getValueType();
1271
1272 // i32 load from lower address.
Justin Lebar9c375812016-07-15 18:27:10 +00001273 SDValue Lo = DAG.getLoad(MVT::i32, DL, Chain, Ptr, MachinePointerInfo(),
1274 Nd.getAlignment(), Nd.getMemOperand()->getFlags());
Akira Hatanaka63791212013-09-07 00:52:30 +00001275
1276 // i32 load from higher address.
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001277 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, DL, PtrVT));
Justin Lebar9c375812016-07-15 18:27:10 +00001278 SDValue Hi = DAG.getLoad(
1279 MVT::i32, DL, Lo.getValue(1), Ptr, MachinePointerInfo(),
1280 std::min(Nd.getAlignment(), 4U), Nd.getMemOperand()->getFlags());
Akira Hatanaka63791212013-09-07 00:52:30 +00001281
Eric Christopher1c29a652014-07-18 22:55:25 +00001282 if (!Subtarget.isLittle())
Akira Hatanaka63791212013-09-07 00:52:30 +00001283 std::swap(Lo, Hi);
1284
1285 SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
1286 SDValue Ops[2] = {BP, Hi.getValue(1)};
Craig Topper64941d92014-04-27 19:20:57 +00001287 return DAG.getMergeValues(Ops, DL);
Akira Hatanaka63791212013-09-07 00:52:30 +00001288}
1289
1290SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1291 StoreSDNode &Nd = *cast<StoreSDNode>(Op);
1292
1293 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1294 return MipsTargetLowering::lowerSTORE(Op, DAG);
1295
1296 // Replace a double precision store with two extractelement64s and i32 stores.
1297 SDLoc DL(Op);
1298 SDValue Val = Nd.getValue(), Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1299 EVT PtrVT = Ptr.getValueType();
1300 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001301 Val, DAG.getConstant(0, DL, MVT::i32));
Akira Hatanaka63791212013-09-07 00:52:30 +00001302 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001303 Val, DAG.getConstant(1, DL, MVT::i32));
Akira Hatanaka63791212013-09-07 00:52:30 +00001304
Eric Christopher1c29a652014-07-18 22:55:25 +00001305 if (!Subtarget.isLittle())
Akira Hatanaka63791212013-09-07 00:52:30 +00001306 std::swap(Lo, Hi);
1307
1308 // i32 store to lower address.
Justin Lebar9c375812016-07-15 18:27:10 +00001309 Chain =
1310 DAG.getStore(Chain, DL, Lo, Ptr, MachinePointerInfo(), Nd.getAlignment(),
1311 Nd.getMemOperand()->getFlags(), Nd.getAAInfo());
Akira Hatanaka63791212013-09-07 00:52:30 +00001312
1313 // i32 store to higher address.
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001314 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, DL, PtrVT));
Akira Hatanaka63791212013-09-07 00:52:30 +00001315 return DAG.getStore(Chain, DL, Hi, Ptr, MachinePointerInfo(),
Justin Lebar9c375812016-07-15 18:27:10 +00001316 std::min(Nd.getAlignment(), 4U),
1317 Nd.getMemOperand()->getFlags(), Nd.getAAInfo());
Akira Hatanaka63791212013-09-07 00:52:30 +00001318}
1319
Akira Hatanakabe8612f2013-03-30 01:36:35 +00001320SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
1321 bool HasLo, bool HasHi,
1322 SelectionDAG &DAG) const {
Daniel Sanders308181e2014-06-12 10:44:10 +00001323 // MIPS32r6/MIPS64r6 removed accumulator based multiplies.
Eric Christopher1c29a652014-07-18 22:55:25 +00001324 assert(!Subtarget.hasMips32r6());
Daniel Sanders308181e2014-06-12 10:44:10 +00001325
Akira Hatanakabe8612f2013-03-30 01:36:35 +00001326 EVT Ty = Op.getOperand(0).getValueType();
Andrew Trickef9de2a2013-05-25 02:42:55 +00001327 SDLoc DL(Op);
Akira Hatanakabe8612f2013-03-30 01:36:35 +00001328 SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
1329 Op.getOperand(0), Op.getOperand(1));
1330 SDValue Lo, Hi;
1331
1332 if (HasLo)
Akira Hatanakad98c99f2013-10-15 01:12:50 +00001333 Lo = DAG.getNode(MipsISD::MFLO, DL, Ty, Mult);
Akira Hatanakabe8612f2013-03-30 01:36:35 +00001334 if (HasHi)
Akira Hatanakad98c99f2013-10-15 01:12:50 +00001335 Hi = DAG.getNode(MipsISD::MFHI, DL, Ty, Mult);
Akira Hatanakabe8612f2013-03-30 01:36:35 +00001336
1337 if (!HasLo || !HasHi)
1338 return HasLo ? Lo : Hi;
1339
1340 SDValue Vals[] = { Lo, Hi };
Craig Topper64941d92014-04-27 19:20:57 +00001341 return DAG.getMergeValues(Vals, DL);
Akira Hatanakabe8612f2013-03-30 01:36:35 +00001342}
1343
Benjamin Kramerbdc49562016-06-12 15:39:02 +00001344static SDValue initAccumulator(SDValue In, const SDLoc &DL, SelectionDAG &DAG) {
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00001345 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001346 DAG.getConstant(0, DL, MVT::i32));
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00001347 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001348 DAG.getConstant(1, DL, MVT::i32));
Akira Hatanakad98c99f2013-10-15 01:12:50 +00001349 return DAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped, InLo, InHi);
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00001350}
1351
Benjamin Kramerbdc49562016-06-12 15:39:02 +00001352static SDValue extractLOHI(SDValue Op, const SDLoc &DL, SelectionDAG &DAG) {
Akira Hatanakad98c99f2013-10-15 01:12:50 +00001353 SDValue Lo = DAG.getNode(MipsISD::MFLO, DL, MVT::i32, Op);
1354 SDValue Hi = DAG.getNode(MipsISD::MFHI, DL, MVT::i32, Op);
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00001355 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
1356}
1357
1358// This function expands mips intrinsic nodes which have 64-bit input operands
1359// or output values.
1360//
1361// out64 = intrinsic-node in64
1362// =>
1363// lo = copy (extract-element (in64, 0))
1364// hi = copy (extract-element (in64, 1))
1365// mips-specific-node
1366// v0 = copy lo
1367// v1 = copy hi
1368// out64 = merge-values (v0, v1)
1369//
1370static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
Andrew Trickef9de2a2013-05-25 02:42:55 +00001371 SDLoc DL(Op);
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00001372 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
1373 SmallVector<SDValue, 3> Ops;
1374 unsigned OpNo = 0;
1375
1376 // See if Op has a chain input.
1377 if (HasChainIn)
1378 Ops.push_back(Op->getOperand(OpNo++));
1379
1380 // The next operand is the intrinsic opcode.
1381 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant);
1382
1383 // See if the next operand has type i64.
1384 SDValue Opnd = Op->getOperand(++OpNo), In64;
1385
1386 if (Opnd.getValueType() == MVT::i64)
1387 In64 = initAccumulator(Opnd, DL, DAG);
1388 else
1389 Ops.push_back(Opnd);
1390
1391 // Push the remaining operands.
1392 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo)
1393 Ops.push_back(Op->getOperand(OpNo));
1394
1395 // Add In64 to the end of the list.
1396 if (In64.getNode())
1397 Ops.push_back(In64);
1398
1399 // Scan output.
1400 SmallVector<EVT, 2> ResTys;
1401
1402 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end();
1403 I != E; ++I)
1404 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I);
1405
1406 // Create node.
Craig Topper48d114b2014-04-26 18:35:24 +00001407 SDValue Val = DAG.getNode(Opc, DL, ResTys, Ops);
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00001408 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val;
1409
1410 if (!HasChainIn)
1411 return Out;
1412
1413 assert(Val->getValueType(1) == MVT::Other);
1414 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) };
Craig Topper64941d92014-04-27 19:20:57 +00001415 return DAG.getMergeValues(Vals, DL);
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00001416}
1417
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +00001418// Lower an MSA copy intrinsic into the specified SelectionDAG node
1419static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1420 SDLoc DL(Op);
1421 SDValue Vec = Op->getOperand(1);
1422 SDValue Idx = Op->getOperand(2);
1423 EVT ResTy = Op->getValueType(0);
1424 EVT EltTy = Vec->getValueType(0).getVectorElementType();
1425
1426 SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
1427 DAG.getValueType(EltTy));
1428
1429 return Result;
1430}
1431
Daniel Sanders50b80412013-11-15 12:56:49 +00001432static SDValue lowerMSASplatZExt(SDValue Op, unsigned OpNr, SelectionDAG &DAG) {
1433 EVT ResVecTy = Op->getValueType(0);
1434 EVT ViaVecTy = ResVecTy;
1435 SDLoc DL(Op);
Daniel Sanders86d0c8d2013-09-23 14:29:55 +00001436
Daniel Sanders50b80412013-11-15 12:56:49 +00001437 // When ResVecTy == MVT::v2i64, LaneA is the upper 32 bits of the lane and
1438 // LaneB is the lower 32-bits. Otherwise LaneA and LaneB are alternating
1439 // lanes.
1440 SDValue LaneA;
1441 SDValue LaneB = Op->getOperand(2);
1442
1443 if (ResVecTy == MVT::v2i64) {
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001444 LaneA = DAG.getConstant(0, DL, MVT::i32);
Daniel Sandersf49dd822013-09-24 13:33:07 +00001445 ViaVecTy = MVT::v4i32;
Daniel Sanders50b80412013-11-15 12:56:49 +00001446 } else
1447 LaneA = LaneB;
Daniel Sanders86d0c8d2013-09-23 14:29:55 +00001448
Daniel Sanders50b80412013-11-15 12:56:49 +00001449 SDValue Ops[16] = { LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB,
1450 LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB };
Daniel Sandersf49dd822013-09-24 13:33:07 +00001451
Ahmed Bougacha128f8732016-04-26 21:15:30 +00001452 SDValue Result = DAG.getBuildVector(
1453 ViaVecTy, DL, makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
Daniel Sanders50b80412013-11-15 12:56:49 +00001454
1455 if (ViaVecTy != ResVecTy)
1456 Result = DAG.getNode(ISD::BITCAST, DL, ResVecTy, Result);
Daniel Sandersf49dd822013-09-24 13:33:07 +00001457
1458 return Result;
1459}
1460
Simon Dardis548a53f2017-01-10 16:40:57 +00001461static SDValue lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG,
1462 bool IsSigned = false) {
1463 return DAG.getConstant(
1464 APInt(Op->getValueType(0).getScalarType().getSizeInBits(),
1465 Op->getConstantOperandVal(ImmOp), IsSigned),
1466 SDLoc(Op), Op->getValueType(0));
Daniel Sanders50b80412013-11-15 12:56:49 +00001467}
1468
1469static SDValue getBuildVectorSplat(EVT VecTy, SDValue SplatValue,
1470 bool BigEndian, SelectionDAG &DAG) {
1471 EVT ViaVecTy = VecTy;
1472 SDValue SplatValueA = SplatValue;
1473 SDValue SplatValueB = SplatValue;
1474 SDLoc DL(SplatValue);
1475
1476 if (VecTy == MVT::v2i64) {
1477 // v2i64 BUILD_VECTOR must be performed via v4i32 so split into i32's.
1478 ViaVecTy = MVT::v4i32;
1479
1480 SplatValueA = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValue);
1481 SplatValueB = DAG.getNode(ISD::SRL, DL, MVT::i64, SplatValue,
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001482 DAG.getConstant(32, DL, MVT::i32));
Daniel Sanders50b80412013-11-15 12:56:49 +00001483 SplatValueB = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValueB);
1484 }
1485
1486 // We currently hold the parts in little endian order. Swap them if
1487 // necessary.
1488 if (BigEndian)
1489 std::swap(SplatValueA, SplatValueB);
1490
1491 SDValue Ops[16] = { SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1492 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1493 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1494 SplatValueA, SplatValueB, SplatValueA, SplatValueB };
1495
Ahmed Bougacha128f8732016-04-26 21:15:30 +00001496 SDValue Result = DAG.getBuildVector(
1497 ViaVecTy, DL, makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
Daniel Sanders50b80412013-11-15 12:56:49 +00001498
1499 if (VecTy != ViaVecTy)
1500 Result = DAG.getNode(ISD::BITCAST, DL, VecTy, Result);
1501
1502 return Result;
Daniel Sandersf5bd9372013-09-11 10:51:30 +00001503}
1504
Daniel Sandersa5bc99f2013-11-12 10:31:49 +00001505static SDValue lowerMSABinaryBitImmIntr(SDValue Op, SelectionDAG &DAG,
1506 unsigned Opc, SDValue Imm,
1507 bool BigEndian) {
1508 EVT VecTy = Op->getValueType(0);
1509 SDValue Exp2Imm;
1510 SDLoc DL(Op);
1511
Daniel Sanders50b80412013-11-15 12:56:49 +00001512 // The DAG Combiner can't constant fold bitcasted vectors yet so we must do it
1513 // here for now.
Daniel Sandersa5bc99f2013-11-12 10:31:49 +00001514 if (VecTy == MVT::v2i64) {
1515 if (ConstantSDNode *CImm = dyn_cast<ConstantSDNode>(Imm)) {
1516 APInt BitImm = APInt(64, 1) << CImm->getAPIntValue();
1517
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001518 SDValue BitImmHiOp = DAG.getConstant(BitImm.lshr(32).trunc(32), DL,
1519 MVT::i32);
1520 SDValue BitImmLoOp = DAG.getConstant(BitImm.trunc(32), DL, MVT::i32);
Daniel Sanders50b80412013-11-15 12:56:49 +00001521
1522 if (BigEndian)
1523 std::swap(BitImmLoOp, BitImmHiOp);
1524
Ahmed Bougacha128f8732016-04-26 21:15:30 +00001525 Exp2Imm = DAG.getNode(
1526 ISD::BITCAST, DL, MVT::v2i64,
1527 DAG.getBuildVector(MVT::v4i32, DL,
1528 {BitImmLoOp, BitImmHiOp, BitImmLoOp, BitImmHiOp}));
Daniel Sandersa5bc99f2013-11-12 10:31:49 +00001529 }
1530 }
1531
Craig Topper062a2ba2014-04-25 05:30:21 +00001532 if (!Exp2Imm.getNode()) {
Daniel Sandersa5bc99f2013-11-12 10:31:49 +00001533 // We couldnt constant fold, do a vector shift instead
Daniel Sanders50b80412013-11-15 12:56:49 +00001534
1535 // Extend i32 to i64 if necessary. Sign or zero extend doesn't matter since
1536 // only values 0-63 are valid.
1537 if (VecTy == MVT::v2i64)
1538 Imm = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Imm);
1539
1540 Exp2Imm = getBuildVectorSplat(VecTy, Imm, BigEndian, DAG);
1541
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001542 Exp2Imm = DAG.getNode(ISD::SHL, DL, VecTy, DAG.getConstant(1, DL, VecTy),
1543 Exp2Imm);
Daniel Sandersa5bc99f2013-11-12 10:31:49 +00001544 }
1545
1546 return DAG.getNode(Opc, DL, VecTy, Op->getOperand(1), Exp2Imm);
1547}
1548
Daniel Sanders3f6eb542013-11-12 10:45:18 +00001549static SDValue lowerMSABitClear(SDValue Op, SelectionDAG &DAG) {
1550 EVT ResTy = Op->getValueType(0);
Daniel Sanders3f6eb542013-11-12 10:45:18 +00001551 SDLoc DL(Op);
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001552 SDValue One = DAG.getConstant(1, DL, ResTy);
Daniel Sanders3f6eb542013-11-12 10:45:18 +00001553 SDValue Bit = DAG.getNode(ISD::SHL, DL, ResTy, One, Op->getOperand(2));
1554
Daniel Sanders71ce0ca2013-11-15 16:02:04 +00001555 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1),
1556 DAG.getNOT(DL, Bit, ResTy));
Daniel Sanders3f6eb542013-11-12 10:45:18 +00001557}
1558
1559static SDValue lowerMSABitClearImm(SDValue Op, SelectionDAG &DAG) {
1560 SDLoc DL(Op);
1561 EVT ResTy = Op->getValueType(0);
Sanjay Patel1ed771f2016-09-14 16:37:15 +00001562 APInt BitImm = APInt(ResTy.getScalarSizeInBits(), 1)
Daniel Sanders50b80412013-11-15 12:56:49 +00001563 << cast<ConstantSDNode>(Op->getOperand(2))->getAPIntValue();
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001564 SDValue BitMask = DAG.getConstant(~BitImm, DL, ResTy);
Daniel Sanders3f6eb542013-11-12 10:45:18 +00001565
1566 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1), BitMask);
1567}
1568
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00001569SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
1570 SelectionDAG &DAG) const {
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001571 SDLoc DL(Op);
Simon Dardis548a53f2017-01-10 16:40:57 +00001572 unsigned Intrinsic = cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue();
1573 switch (Intrinsic) {
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00001574 default:
1575 return SDValue();
1576 case Intrinsic::mips_shilo:
1577 return lowerDSPIntr(Op, DAG, MipsISD::SHILO);
1578 case Intrinsic::mips_dpau_h_qbl:
1579 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL);
1580 case Intrinsic::mips_dpau_h_qbr:
1581 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR);
1582 case Intrinsic::mips_dpsu_h_qbl:
1583 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL);
1584 case Intrinsic::mips_dpsu_h_qbr:
1585 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR);
1586 case Intrinsic::mips_dpa_w_ph:
1587 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH);
1588 case Intrinsic::mips_dps_w_ph:
1589 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH);
1590 case Intrinsic::mips_dpax_w_ph:
1591 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH);
1592 case Intrinsic::mips_dpsx_w_ph:
1593 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH);
1594 case Intrinsic::mips_mulsa_w_ph:
1595 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH);
1596 case Intrinsic::mips_mult:
1597 return lowerDSPIntr(Op, DAG, MipsISD::Mult);
1598 case Intrinsic::mips_multu:
1599 return lowerDSPIntr(Op, DAG, MipsISD::Multu);
1600 case Intrinsic::mips_madd:
1601 return lowerDSPIntr(Op, DAG, MipsISD::MAdd);
1602 case Intrinsic::mips_maddu:
1603 return lowerDSPIntr(Op, DAG, MipsISD::MAddu);
1604 case Intrinsic::mips_msub:
1605 return lowerDSPIntr(Op, DAG, MipsISD::MSub);
1606 case Intrinsic::mips_msubu:
1607 return lowerDSPIntr(Op, DAG, MipsISD::MSubu);
Daniel Sandersfa5ab1c2013-09-11 10:28:16 +00001608 case Intrinsic::mips_addv_b:
1609 case Intrinsic::mips_addv_h:
1610 case Intrinsic::mips_addv_w:
1611 case Intrinsic::mips_addv_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001612 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1613 Op->getOperand(2));
Daniel Sanders86d0c8d2013-09-23 14:29:55 +00001614 case Intrinsic::mips_addvi_b:
1615 case Intrinsic::mips_addvi_h:
1616 case Intrinsic::mips_addvi_w:
1617 case Intrinsic::mips_addvi_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001618 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1619 lowerMSASplatImm(Op, 2, DAG));
Daniel Sanders8ca81e42013-09-23 12:57:42 +00001620 case Intrinsic::mips_and_v:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001621 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1622 Op->getOperand(2));
Daniel Sandersbfc39ce2013-09-24 12:32:47 +00001623 case Intrinsic::mips_andi_b:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001624 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1625 lowerMSASplatImm(Op, 2, DAG));
Daniel Sanders3f6eb542013-11-12 10:45:18 +00001626 case Intrinsic::mips_bclr_b:
1627 case Intrinsic::mips_bclr_h:
1628 case Intrinsic::mips_bclr_w:
1629 case Intrinsic::mips_bclr_d:
1630 return lowerMSABitClear(Op, DAG);
1631 case Intrinsic::mips_bclri_b:
1632 case Intrinsic::mips_bclri_h:
1633 case Intrinsic::mips_bclri_w:
1634 case Intrinsic::mips_bclri_d:
1635 return lowerMSABitClearImm(Op, DAG);
Daniel Sandersd74b1302013-10-30 14:45:14 +00001636 case Intrinsic::mips_binsli_b:
1637 case Intrinsic::mips_binsli_h:
1638 case Intrinsic::mips_binsli_w:
1639 case Intrinsic::mips_binsli_d: {
Daniel Sandersdf2215452014-03-12 11:54:00 +00001640 // binsli_x(IfClear, IfSet, nbits) -> (vselect LBitsMask, IfSet, IfClear)
Daniel Sandersd74b1302013-10-30 14:45:14 +00001641 EVT VecTy = Op->getValueType(0);
1642 EVT EltTy = VecTy.getVectorElementType();
Simon Dardis548a53f2017-01-10 16:40:57 +00001643 if (Op->getConstantOperandVal(3) >= EltTy.getSizeInBits())
1644 report_fatal_error("Immediate out of range");
Daniel Sandersd74b1302013-10-30 14:45:14 +00001645 APInt Mask = APInt::getHighBitsSet(EltTy.getSizeInBits(),
1646 Op->getConstantOperandVal(3));
1647 return DAG.getNode(ISD::VSELECT, DL, VecTy,
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001648 DAG.getConstant(Mask, DL, VecTy, true),
1649 Op->getOperand(2), Op->getOperand(1));
Daniel Sandersd74b1302013-10-30 14:45:14 +00001650 }
1651 case Intrinsic::mips_binsri_b:
1652 case Intrinsic::mips_binsri_h:
1653 case Intrinsic::mips_binsri_w:
1654 case Intrinsic::mips_binsri_d: {
Daniel Sandersdf2215452014-03-12 11:54:00 +00001655 // binsri_x(IfClear, IfSet, nbits) -> (vselect RBitsMask, IfSet, IfClear)
Daniel Sandersd74b1302013-10-30 14:45:14 +00001656 EVT VecTy = Op->getValueType(0);
1657 EVT EltTy = VecTy.getVectorElementType();
Simon Dardis548a53f2017-01-10 16:40:57 +00001658 if (Op->getConstantOperandVal(3) >= EltTy.getSizeInBits())
1659 report_fatal_error("Immediate out of range");
Daniel Sandersd74b1302013-10-30 14:45:14 +00001660 APInt Mask = APInt::getLowBitsSet(EltTy.getSizeInBits(),
1661 Op->getConstantOperandVal(3));
1662 return DAG.getNode(ISD::VSELECT, DL, VecTy,
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001663 DAG.getConstant(Mask, DL, VecTy, true),
1664 Op->getOperand(2), Op->getOperand(1));
Daniel Sandersd74b1302013-10-30 14:45:14 +00001665 }
Daniel Sandersab94b532013-10-30 15:20:38 +00001666 case Intrinsic::mips_bmnz_v:
1667 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1668 Op->getOperand(2), Op->getOperand(1));
1669 case Intrinsic::mips_bmnzi_b:
1670 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1671 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(2),
1672 Op->getOperand(1));
1673 case Intrinsic::mips_bmz_v:
1674 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1675 Op->getOperand(1), Op->getOperand(2));
1676 case Intrinsic::mips_bmzi_b:
1677 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1678 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(1),
1679 Op->getOperand(2));
Daniel Sandersa5bc99f2013-11-12 10:31:49 +00001680 case Intrinsic::mips_bneg_b:
1681 case Intrinsic::mips_bneg_h:
1682 case Intrinsic::mips_bneg_w:
1683 case Intrinsic::mips_bneg_d: {
1684 EVT VecTy = Op->getValueType(0);
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001685 SDValue One = DAG.getConstant(1, DL, VecTy);
Daniel Sandersa5bc99f2013-11-12 10:31:49 +00001686
1687 return DAG.getNode(ISD::XOR, DL, VecTy, Op->getOperand(1),
1688 DAG.getNode(ISD::SHL, DL, VecTy, One,
1689 Op->getOperand(2)));
1690 }
1691 case Intrinsic::mips_bnegi_b:
1692 case Intrinsic::mips_bnegi_h:
1693 case Intrinsic::mips_bnegi_w:
1694 case Intrinsic::mips_bnegi_d:
1695 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::XOR, Op->getOperand(2),
Eric Christopher1c29a652014-07-18 22:55:25 +00001696 !Subtarget.isLittle());
Daniel Sandersce09d072013-08-28 12:14:50 +00001697 case Intrinsic::mips_bnz_b:
1698 case Intrinsic::mips_bnz_h:
1699 case Intrinsic::mips_bnz_w:
1700 case Intrinsic::mips_bnz_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001701 return DAG.getNode(MipsISD::VALL_NONZERO, DL, Op->getValueType(0),
1702 Op->getOperand(1));
Daniel Sandersce09d072013-08-28 12:14:50 +00001703 case Intrinsic::mips_bnz_v:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001704 return DAG.getNode(MipsISD::VANY_NONZERO, DL, Op->getValueType(0),
1705 Op->getOperand(1));
Daniel Sanderse1d24352013-09-24 12:04:44 +00001706 case Intrinsic::mips_bsel_v:
Daniel Sandersdf2215452014-03-12 11:54:00 +00001707 // bsel_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001708 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
Daniel Sandersdf2215452014-03-12 11:54:00 +00001709 Op->getOperand(1), Op->getOperand(3),
1710 Op->getOperand(2));
Daniel Sanderse1d24352013-09-24 12:04:44 +00001711 case Intrinsic::mips_bseli_b:
Daniel Sandersdf2215452014-03-12 11:54:00 +00001712 // bseli_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001713 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
Daniel Sandersdf2215452014-03-12 11:54:00 +00001714 Op->getOperand(1), lowerMSASplatImm(Op, 3, DAG),
1715 Op->getOperand(2));
Daniel Sandersa5bc99f2013-11-12 10:31:49 +00001716 case Intrinsic::mips_bset_b:
1717 case Intrinsic::mips_bset_h:
1718 case Intrinsic::mips_bset_w:
1719 case Intrinsic::mips_bset_d: {
1720 EVT VecTy = Op->getValueType(0);
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00001721 SDValue One = DAG.getConstant(1, DL, VecTy);
Daniel Sandersa5bc99f2013-11-12 10:31:49 +00001722
1723 return DAG.getNode(ISD::OR, DL, VecTy, Op->getOperand(1),
1724 DAG.getNode(ISD::SHL, DL, VecTy, One,
1725 Op->getOperand(2)));
1726 }
1727 case Intrinsic::mips_bseti_b:
1728 case Intrinsic::mips_bseti_h:
1729 case Intrinsic::mips_bseti_w:
1730 case Intrinsic::mips_bseti_d:
1731 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::OR, Op->getOperand(2),
Eric Christopher1c29a652014-07-18 22:55:25 +00001732 !Subtarget.isLittle());
Daniel Sandersce09d072013-08-28 12:14:50 +00001733 case Intrinsic::mips_bz_b:
1734 case Intrinsic::mips_bz_h:
1735 case Intrinsic::mips_bz_w:
1736 case Intrinsic::mips_bz_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001737 return DAG.getNode(MipsISD::VALL_ZERO, DL, Op->getValueType(0),
1738 Op->getOperand(1));
Daniel Sandersce09d072013-08-28 12:14:50 +00001739 case Intrinsic::mips_bz_v:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001740 return DAG.getNode(MipsISD::VANY_ZERO, DL, Op->getValueType(0),
1741 Op->getOperand(1));
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001742 case Intrinsic::mips_ceq_b:
1743 case Intrinsic::mips_ceq_h:
1744 case Intrinsic::mips_ceq_w:
1745 case Intrinsic::mips_ceq_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001746 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001747 Op->getOperand(2), ISD::SETEQ);
1748 case Intrinsic::mips_ceqi_b:
1749 case Intrinsic::mips_ceqi_h:
1750 case Intrinsic::mips_ceqi_w:
1751 case Intrinsic::mips_ceqi_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001752 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Simon Dardis548a53f2017-01-10 16:40:57 +00001753 lowerMSASplatImm(Op, 2, DAG, true), ISD::SETEQ);
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001754 case Intrinsic::mips_cle_s_b:
1755 case Intrinsic::mips_cle_s_h:
1756 case Intrinsic::mips_cle_s_w:
1757 case Intrinsic::mips_cle_s_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001758 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001759 Op->getOperand(2), ISD::SETLE);
1760 case Intrinsic::mips_clei_s_b:
1761 case Intrinsic::mips_clei_s_h:
1762 case Intrinsic::mips_clei_s_w:
1763 case Intrinsic::mips_clei_s_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001764 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Simon Dardis548a53f2017-01-10 16:40:57 +00001765 lowerMSASplatImm(Op, 2, DAG, true), ISD::SETLE);
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001766 case Intrinsic::mips_cle_u_b:
1767 case Intrinsic::mips_cle_u_h:
1768 case Intrinsic::mips_cle_u_w:
1769 case Intrinsic::mips_cle_u_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001770 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001771 Op->getOperand(2), ISD::SETULE);
1772 case Intrinsic::mips_clei_u_b:
1773 case Intrinsic::mips_clei_u_h:
1774 case Intrinsic::mips_clei_u_w:
1775 case Intrinsic::mips_clei_u_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001776 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001777 lowerMSASplatImm(Op, 2, DAG), ISD::SETULE);
1778 case Intrinsic::mips_clt_s_b:
1779 case Intrinsic::mips_clt_s_h:
1780 case Intrinsic::mips_clt_s_w:
1781 case Intrinsic::mips_clt_s_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001782 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001783 Op->getOperand(2), ISD::SETLT);
1784 case Intrinsic::mips_clti_s_b:
1785 case Intrinsic::mips_clti_s_h:
1786 case Intrinsic::mips_clti_s_w:
1787 case Intrinsic::mips_clti_s_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001788 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Simon Dardis548a53f2017-01-10 16:40:57 +00001789 lowerMSASplatImm(Op, 2, DAG, true), ISD::SETLT);
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001790 case Intrinsic::mips_clt_u_b:
1791 case Intrinsic::mips_clt_u_h:
1792 case Intrinsic::mips_clt_u_w:
1793 case Intrinsic::mips_clt_u_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001794 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001795 Op->getOperand(2), ISD::SETULT);
1796 case Intrinsic::mips_clti_u_b:
1797 case Intrinsic::mips_clti_u_h:
1798 case Intrinsic::mips_clti_u_w:
1799 case Intrinsic::mips_clti_u_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001800 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001801 lowerMSASplatImm(Op, 2, DAG), ISD::SETULT);
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +00001802 case Intrinsic::mips_copy_s_b:
1803 case Intrinsic::mips_copy_s_h:
1804 case Intrinsic::mips_copy_s_w:
1805 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
Daniel Sanders7f3d9462013-09-27 13:04:21 +00001806 case Intrinsic::mips_copy_s_d:
Eric Christopher1c29a652014-07-18 22:55:25 +00001807 if (Subtarget.hasMips64())
Matheus Almeida74070322014-01-29 14:05:28 +00001808 // Lower directly into VEXTRACT_SEXT_ELT since i64 is legal on Mips64.
1809 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1810 else {
1811 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1812 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1813 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1814 Op->getValueType(0), Op->getOperand(1),
1815 Op->getOperand(2));
1816 }
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +00001817 case Intrinsic::mips_copy_u_b:
1818 case Intrinsic::mips_copy_u_h:
1819 case Intrinsic::mips_copy_u_w:
1820 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
Daniel Sanders7f3d9462013-09-27 13:04:21 +00001821 case Intrinsic::mips_copy_u_d:
Eric Christopher1c29a652014-07-18 22:55:25 +00001822 if (Subtarget.hasMips64())
Matheus Almeida74070322014-01-29 14:05:28 +00001823 // Lower directly into VEXTRACT_ZEXT_ELT since i64 is legal on Mips64.
1824 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1825 else {
1826 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1827 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1828 // Note: When i64 is illegal, this results in copy_s.w instructions
1829 // instead of copy_u.w instructions. This makes no difference to the
1830 // behaviour since i64 is only illegal when the register file is 32-bit.
1831 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1832 Op->getValueType(0), Op->getOperand(1),
1833 Op->getOperand(2));
1834 }
Daniel Sanders607952b2013-09-11 10:38:58 +00001835 case Intrinsic::mips_div_s_b:
1836 case Intrinsic::mips_div_s_h:
1837 case Intrinsic::mips_div_s_w:
1838 case Intrinsic::mips_div_s_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001839 return DAG.getNode(ISD::SDIV, DL, Op->getValueType(0), Op->getOperand(1),
1840 Op->getOperand(2));
Daniel Sanders607952b2013-09-11 10:38:58 +00001841 case Intrinsic::mips_div_u_b:
1842 case Intrinsic::mips_div_u_h:
1843 case Intrinsic::mips_div_u_w:
1844 case Intrinsic::mips_div_u_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001845 return DAG.getNode(ISD::UDIV, DL, Op->getValueType(0), Op->getOperand(1),
1846 Op->getOperand(2));
Daniel Sandersf5bd9372013-09-11 10:51:30 +00001847 case Intrinsic::mips_fadd_w:
Sanjay Patela2607012015-09-16 16:31:21 +00001848 case Intrinsic::mips_fadd_d: {
1849 // TODO: If intrinsics have fast-math-flags, propagate them.
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001850 return DAG.getNode(ISD::FADD, DL, Op->getValueType(0), Op->getOperand(1),
1851 Op->getOperand(2));
Sanjay Patela2607012015-09-16 16:31:21 +00001852 }
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001853 // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away
1854 case Intrinsic::mips_fceq_w:
1855 case Intrinsic::mips_fceq_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001856 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001857 Op->getOperand(2), ISD::SETOEQ);
1858 case Intrinsic::mips_fcle_w:
1859 case Intrinsic::mips_fcle_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001860 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001861 Op->getOperand(2), ISD::SETOLE);
1862 case Intrinsic::mips_fclt_w:
1863 case Intrinsic::mips_fclt_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001864 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001865 Op->getOperand(2), ISD::SETOLT);
1866 case Intrinsic::mips_fcne_w:
1867 case Intrinsic::mips_fcne_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001868 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001869 Op->getOperand(2), ISD::SETONE);
1870 case Intrinsic::mips_fcor_w:
1871 case Intrinsic::mips_fcor_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001872 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001873 Op->getOperand(2), ISD::SETO);
1874 case Intrinsic::mips_fcueq_w:
1875 case Intrinsic::mips_fcueq_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001876 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001877 Op->getOperand(2), ISD::SETUEQ);
1878 case Intrinsic::mips_fcule_w:
1879 case Intrinsic::mips_fcule_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001880 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001881 Op->getOperand(2), ISD::SETULE);
1882 case Intrinsic::mips_fcult_w:
1883 case Intrinsic::mips_fcult_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001884 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001885 Op->getOperand(2), ISD::SETULT);
1886 case Intrinsic::mips_fcun_w:
1887 case Intrinsic::mips_fcun_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001888 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001889 Op->getOperand(2), ISD::SETUO);
1890 case Intrinsic::mips_fcune_w:
1891 case Intrinsic::mips_fcune_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001892 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
Daniel Sandersfd538dc2013-09-24 10:46:19 +00001893 Op->getOperand(2), ISD::SETUNE);
Daniel Sandersf5bd9372013-09-11 10:51:30 +00001894 case Intrinsic::mips_fdiv_w:
Sanjay Patela2607012015-09-16 16:31:21 +00001895 case Intrinsic::mips_fdiv_d: {
1896 // TODO: If intrinsics have fast-math-flags, propagate them.
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001897 return DAG.getNode(ISD::FDIV, DL, Op->getValueType(0), Op->getOperand(1),
1898 Op->getOperand(2));
Sanjay Patela2607012015-09-16 16:31:21 +00001899 }
Daniel Sanders015972b2013-10-11 10:00:06 +00001900 case Intrinsic::mips_ffint_u_w:
1901 case Intrinsic::mips_ffint_u_d:
1902 return DAG.getNode(ISD::UINT_TO_FP, DL, Op->getValueType(0),
1903 Op->getOperand(1));
1904 case Intrinsic::mips_ffint_s_w:
1905 case Intrinsic::mips_ffint_s_d:
1906 return DAG.getNode(ISD::SINT_TO_FP, DL, Op->getValueType(0),
1907 Op->getOperand(1));
Daniel Sanders7a289d02013-09-23 12:02:46 +00001908 case Intrinsic::mips_fill_b:
1909 case Intrinsic::mips_fill_h:
Daniel Sandersc72593e2013-09-27 13:20:41 +00001910 case Intrinsic::mips_fill_w:
1911 case Intrinsic::mips_fill_d: {
Daniel Sandersf49dd822013-09-24 13:33:07 +00001912 EVT ResTy = Op->getValueType(0);
Benjamin Kramer6cd780f2015-02-17 15:29:18 +00001913 SmallVector<SDValue, 16> Ops(ResTy.getVectorNumElements(),
1914 Op->getOperand(1));
Daniel Sandersf49dd822013-09-24 13:33:07 +00001915
Daniel Sandersc72593e2013-09-27 13:20:41 +00001916 // If ResTy is v2i64 then the type legalizer will break this node down into
1917 // an equivalent v4i32.
Ahmed Bougacha128f8732016-04-26 21:15:30 +00001918 return DAG.getBuildVector(ResTy, DL, Ops);
Daniel Sandersf49dd822013-09-24 13:33:07 +00001919 }
Daniel Sandersa9521602013-10-23 10:36:52 +00001920 case Intrinsic::mips_fexp2_w:
1921 case Intrinsic::mips_fexp2_d: {
Sanjay Patela2607012015-09-16 16:31:21 +00001922 // TODO: If intrinsics have fast-math-flags, propagate them.
Daniel Sandersa9521602013-10-23 10:36:52 +00001923 EVT ResTy = Op->getValueType(0);
1924 return DAG.getNode(
1925 ISD::FMUL, SDLoc(Op), ResTy, Op->getOperand(1),
1926 DAG.getNode(ISD::FEXP2, SDLoc(Op), ResTy, Op->getOperand(2)));
1927 }
Daniel Sandersf5bd9372013-09-11 10:51:30 +00001928 case Intrinsic::mips_flog2_w:
1929 case Intrinsic::mips_flog2_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001930 return DAG.getNode(ISD::FLOG2, DL, Op->getValueType(0), Op->getOperand(1));
Daniel Sandersd7103f32013-10-11 10:14:25 +00001931 case Intrinsic::mips_fmadd_w:
1932 case Intrinsic::mips_fmadd_d:
1933 return DAG.getNode(ISD::FMA, SDLoc(Op), Op->getValueType(0),
1934 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
Daniel Sandersf5bd9372013-09-11 10:51:30 +00001935 case Intrinsic::mips_fmul_w:
Sanjay Patela2607012015-09-16 16:31:21 +00001936 case Intrinsic::mips_fmul_d: {
1937 // TODO: If intrinsics have fast-math-flags, propagate them.
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001938 return DAG.getNode(ISD::FMUL, DL, Op->getValueType(0), Op->getOperand(1),
1939 Op->getOperand(2));
Sanjay Patela2607012015-09-16 16:31:21 +00001940 }
Daniel Sanderse67bd872013-10-11 10:27:32 +00001941 case Intrinsic::mips_fmsub_w:
1942 case Intrinsic::mips_fmsub_d: {
Sanjay Patela2607012015-09-16 16:31:21 +00001943 // TODO: If intrinsics have fast-math-flags, propagate them.
Daniel Sanderse67bd872013-10-11 10:27:32 +00001944 EVT ResTy = Op->getValueType(0);
1945 return DAG.getNode(ISD::FSUB, SDLoc(Op), ResTy, Op->getOperand(1),
1946 DAG.getNode(ISD::FMUL, SDLoc(Op), ResTy,
1947 Op->getOperand(2), Op->getOperand(3)));
1948 }
Daniel Sandersf5bd9372013-09-11 10:51:30 +00001949 case Intrinsic::mips_frint_w:
1950 case Intrinsic::mips_frint_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001951 return DAG.getNode(ISD::FRINT, DL, Op->getValueType(0), Op->getOperand(1));
Daniel Sandersf5bd9372013-09-11 10:51:30 +00001952 case Intrinsic::mips_fsqrt_w:
1953 case Intrinsic::mips_fsqrt_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001954 return DAG.getNode(ISD::FSQRT, DL, Op->getValueType(0), Op->getOperand(1));
Daniel Sandersf5bd9372013-09-11 10:51:30 +00001955 case Intrinsic::mips_fsub_w:
Sanjay Patela2607012015-09-16 16:31:21 +00001956 case Intrinsic::mips_fsub_d: {
1957 // TODO: If intrinsics have fast-math-flags, propagate them.
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001958 return DAG.getNode(ISD::FSUB, DL, Op->getValueType(0), Op->getOperand(1),
1959 Op->getOperand(2));
Sanjay Patela2607012015-09-16 16:31:21 +00001960 }
Daniel Sanders015972b2013-10-11 10:00:06 +00001961 case Intrinsic::mips_ftrunc_u_w:
1962 case Intrinsic::mips_ftrunc_u_d:
1963 return DAG.getNode(ISD::FP_TO_UINT, DL, Op->getValueType(0),
1964 Op->getOperand(1));
1965 case Intrinsic::mips_ftrunc_s_w:
1966 case Intrinsic::mips_ftrunc_s_d:
1967 return DAG.getNode(ISD::FP_TO_SINT, DL, Op->getValueType(0),
1968 Op->getOperand(1));
Daniel Sanders2ed228b2013-09-24 14:36:12 +00001969 case Intrinsic::mips_ilvev_b:
1970 case Intrinsic::mips_ilvev_h:
1971 case Intrinsic::mips_ilvev_w:
1972 case Intrinsic::mips_ilvev_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001973 return DAG.getNode(MipsISD::ILVEV, DL, Op->getValueType(0),
Daniel Sanders2ed228b2013-09-24 14:36:12 +00001974 Op->getOperand(1), Op->getOperand(2));
1975 case Intrinsic::mips_ilvl_b:
1976 case Intrinsic::mips_ilvl_h:
1977 case Intrinsic::mips_ilvl_w:
1978 case Intrinsic::mips_ilvl_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001979 return DAG.getNode(MipsISD::ILVL, DL, Op->getValueType(0),
Daniel Sanders2ed228b2013-09-24 14:36:12 +00001980 Op->getOperand(1), Op->getOperand(2));
1981 case Intrinsic::mips_ilvod_b:
1982 case Intrinsic::mips_ilvod_h:
1983 case Intrinsic::mips_ilvod_w:
1984 case Intrinsic::mips_ilvod_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001985 return DAG.getNode(MipsISD::ILVOD, DL, Op->getValueType(0),
Daniel Sanders2ed228b2013-09-24 14:36:12 +00001986 Op->getOperand(1), Op->getOperand(2));
1987 case Intrinsic::mips_ilvr_b:
1988 case Intrinsic::mips_ilvr_h:
1989 case Intrinsic::mips_ilvr_w:
1990 case Intrinsic::mips_ilvr_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00001991 return DAG.getNode(MipsISD::ILVR, DL, Op->getValueType(0),
Daniel Sanders2ed228b2013-09-24 14:36:12 +00001992 Op->getOperand(1), Op->getOperand(2));
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +00001993 case Intrinsic::mips_insert_b:
1994 case Intrinsic::mips_insert_h:
1995 case Intrinsic::mips_insert_w:
Daniel Sanders6098b332013-09-27 13:36:54 +00001996 case Intrinsic::mips_insert_d:
1997 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(Op), Op->getValueType(0),
1998 Op->getOperand(1), Op->getOperand(3), Op->getOperand(2));
Daniel Sandersb50ccf82014-04-01 10:35:28 +00001999 case Intrinsic::mips_insve_b:
2000 case Intrinsic::mips_insve_h:
2001 case Intrinsic::mips_insve_w:
Simon Dardis548a53f2017-01-10 16:40:57 +00002002 case Intrinsic::mips_insve_d: {
2003 // Report an error for out of range values.
2004 int64_t Max;
2005 switch (Intrinsic) {
2006 case Intrinsic::mips_insve_b: Max = 15; break;
2007 case Intrinsic::mips_insve_h: Max = 7; break;
2008 case Intrinsic::mips_insve_w: Max = 3; break;
2009 case Intrinsic::mips_insve_d: Max = 1; break;
2010 default: llvm_unreachable("Unmatched intrinsic");
2011 }
2012 int64_t Value = cast<ConstantSDNode>(Op->getOperand(2))->getSExtValue();
2013 if (Value < 0 || Value > Max)
2014 report_fatal_error("Immediate out of range");
Daniel Sandersb50ccf82014-04-01 10:35:28 +00002015 return DAG.getNode(MipsISD::INSVE, DL, Op->getValueType(0),
2016 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3),
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00002017 DAG.getConstant(0, DL, MVT::i32));
Simon Dardis548a53f2017-01-10 16:40:57 +00002018 }
Daniel Sanders7a289d02013-09-23 12:02:46 +00002019 case Intrinsic::mips_ldi_b:
2020 case Intrinsic::mips_ldi_h:
2021 case Intrinsic::mips_ldi_w:
2022 case Intrinsic::mips_ldi_d:
Simon Dardis548a53f2017-01-10 16:40:57 +00002023 return lowerMSASplatImm(Op, 1, DAG, true);
Matheus Almeida4b27eb52014-02-10 12:05:17 +00002024 case Intrinsic::mips_lsa:
2025 case Intrinsic::mips_dlsa: {
Daniel Sandersa4eaf592013-10-17 13:38:20 +00002026 EVT ResTy = Op->getValueType(0);
2027 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
2028 DAG.getNode(ISD::SHL, SDLoc(Op), ResTy,
2029 Op->getOperand(2), Op->getOperand(3)));
2030 }
Daniel Sanders50e5ed32013-10-11 10:50:42 +00002031 case Intrinsic::mips_maddv_b:
2032 case Intrinsic::mips_maddv_h:
2033 case Intrinsic::mips_maddv_w:
2034 case Intrinsic::mips_maddv_d: {
2035 EVT ResTy = Op->getValueType(0);
2036 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
2037 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
2038 Op->getOperand(2), Op->getOperand(3)));
2039 }
Daniel Sanders3ce56622013-09-24 12:18:31 +00002040 case Intrinsic::mips_max_s_b:
2041 case Intrinsic::mips_max_s_h:
2042 case Intrinsic::mips_max_s_w:
2043 case Intrinsic::mips_max_s_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002044 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
2045 Op->getOperand(1), Op->getOperand(2));
Daniel Sanders3ce56622013-09-24 12:18:31 +00002046 case Intrinsic::mips_max_u_b:
2047 case Intrinsic::mips_max_u_h:
2048 case Intrinsic::mips_max_u_w:
2049 case Intrinsic::mips_max_u_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002050 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
2051 Op->getOperand(1), Op->getOperand(2));
Daniel Sanders3ce56622013-09-24 12:18:31 +00002052 case Intrinsic::mips_maxi_s_b:
2053 case Intrinsic::mips_maxi_s_h:
2054 case Intrinsic::mips_maxi_s_w:
2055 case Intrinsic::mips_maxi_s_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002056 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
Simon Dardis548a53f2017-01-10 16:40:57 +00002057 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG, true));
Daniel Sanders3ce56622013-09-24 12:18:31 +00002058 case Intrinsic::mips_maxi_u_b:
2059 case Intrinsic::mips_maxi_u_h:
2060 case Intrinsic::mips_maxi_u_w:
2061 case Intrinsic::mips_maxi_u_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002062 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
2063 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
Daniel Sanders3ce56622013-09-24 12:18:31 +00002064 case Intrinsic::mips_min_s_b:
2065 case Intrinsic::mips_min_s_h:
2066 case Intrinsic::mips_min_s_w:
2067 case Intrinsic::mips_min_s_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002068 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
2069 Op->getOperand(1), Op->getOperand(2));
Daniel Sanders3ce56622013-09-24 12:18:31 +00002070 case Intrinsic::mips_min_u_b:
2071 case Intrinsic::mips_min_u_h:
2072 case Intrinsic::mips_min_u_w:
2073 case Intrinsic::mips_min_u_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002074 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
2075 Op->getOperand(1), Op->getOperand(2));
Daniel Sanders3ce56622013-09-24 12:18:31 +00002076 case Intrinsic::mips_mini_s_b:
2077 case Intrinsic::mips_mini_s_h:
2078 case Intrinsic::mips_mini_s_w:
2079 case Intrinsic::mips_mini_s_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002080 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
Simon Dardis548a53f2017-01-10 16:40:57 +00002081 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG, true));
Daniel Sanders3ce56622013-09-24 12:18:31 +00002082 case Intrinsic::mips_mini_u_b:
2083 case Intrinsic::mips_mini_u_h:
2084 case Intrinsic::mips_mini_u_w:
2085 case Intrinsic::mips_mini_u_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002086 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
2087 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
Daniel Sanders0210dd42013-10-01 10:22:35 +00002088 case Intrinsic::mips_mod_s_b:
2089 case Intrinsic::mips_mod_s_h:
2090 case Intrinsic::mips_mod_s_w:
2091 case Intrinsic::mips_mod_s_d:
2092 return DAG.getNode(ISD::SREM, DL, Op->getValueType(0), Op->getOperand(1),
2093 Op->getOperand(2));
2094 case Intrinsic::mips_mod_u_b:
2095 case Intrinsic::mips_mod_u_h:
2096 case Intrinsic::mips_mod_u_w:
2097 case Intrinsic::mips_mod_u_d:
2098 return DAG.getNode(ISD::UREM, DL, Op->getValueType(0), Op->getOperand(1),
2099 Op->getOperand(2));
Daniel Sandersfbcb5822013-09-11 11:58:30 +00002100 case Intrinsic::mips_mulv_b:
2101 case Intrinsic::mips_mulv_h:
2102 case Intrinsic::mips_mulv_w:
2103 case Intrinsic::mips_mulv_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002104 return DAG.getNode(ISD::MUL, DL, Op->getValueType(0), Op->getOperand(1),
2105 Op->getOperand(2));
Daniel Sanders50e5ed32013-10-11 10:50:42 +00002106 case Intrinsic::mips_msubv_b:
2107 case Intrinsic::mips_msubv_h:
2108 case Intrinsic::mips_msubv_w:
2109 case Intrinsic::mips_msubv_d: {
2110 EVT ResTy = Op->getValueType(0);
2111 return DAG.getNode(ISD::SUB, SDLoc(Op), ResTy, Op->getOperand(1),
2112 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
2113 Op->getOperand(2), Op->getOperand(3)));
2114 }
Daniel Sandersfbcb5822013-09-11 11:58:30 +00002115 case Intrinsic::mips_nlzc_b:
2116 case Intrinsic::mips_nlzc_h:
2117 case Intrinsic::mips_nlzc_w:
2118 case Intrinsic::mips_nlzc_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002119 return DAG.getNode(ISD::CTLZ, DL, Op->getValueType(0), Op->getOperand(1));
Daniel Sandersf7456c72013-09-23 13:22:24 +00002120 case Intrinsic::mips_nor_v: {
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002121 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2122 Op->getOperand(1), Op->getOperand(2));
2123 return DAG.getNOT(DL, Res, Res->getValueType(0));
Daniel Sandersf7456c72013-09-23 13:22:24 +00002124 }
Daniel Sandersbfc39ce2013-09-24 12:32:47 +00002125 case Intrinsic::mips_nori_b: {
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002126 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2127 Op->getOperand(1),
2128 lowerMSASplatImm(Op, 2, DAG));
2129 return DAG.getNOT(DL, Res, Res->getValueType(0));
Daniel Sandersbfc39ce2013-09-24 12:32:47 +00002130 }
Daniel Sanders8ca81e42013-09-23 12:57:42 +00002131 case Intrinsic::mips_or_v:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002132 return DAG.getNode(ISD::OR, DL, Op->getValueType(0), Op->getOperand(1),
2133 Op->getOperand(2));
Daniel Sandersbfc39ce2013-09-24 12:32:47 +00002134 case Intrinsic::mips_ori_b:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002135 return DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2136 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002137 case Intrinsic::mips_pckev_b:
2138 case Intrinsic::mips_pckev_h:
2139 case Intrinsic::mips_pckev_w:
2140 case Intrinsic::mips_pckev_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002141 return DAG.getNode(MipsISD::PCKEV, DL, Op->getValueType(0),
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002142 Op->getOperand(1), Op->getOperand(2));
2143 case Intrinsic::mips_pckod_b:
2144 case Intrinsic::mips_pckod_h:
2145 case Intrinsic::mips_pckod_w:
2146 case Intrinsic::mips_pckod_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002147 return DAG.getNode(MipsISD::PCKOD, DL, Op->getValueType(0),
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002148 Op->getOperand(1), Op->getOperand(2));
Daniel Sanders766cb692013-09-23 13:40:21 +00002149 case Intrinsic::mips_pcnt_b:
2150 case Intrinsic::mips_pcnt_h:
2151 case Intrinsic::mips_pcnt_w:
2152 case Intrinsic::mips_pcnt_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002153 return DAG.getNode(ISD::CTPOP, DL, Op->getValueType(0), Op->getOperand(1));
Simon Dardis548a53f2017-01-10 16:40:57 +00002154 case Intrinsic::mips_sat_s_b:
2155 case Intrinsic::mips_sat_s_h:
2156 case Intrinsic::mips_sat_s_w:
2157 case Intrinsic::mips_sat_s_d:
2158 case Intrinsic::mips_sat_u_b:
2159 case Intrinsic::mips_sat_u_h:
2160 case Intrinsic::mips_sat_u_w:
2161 case Intrinsic::mips_sat_u_d: {
2162 // Report an error for out of range values.
2163 int64_t Max;
2164 switch (Intrinsic) {
2165 case Intrinsic::mips_sat_s_b:
2166 case Intrinsic::mips_sat_u_b: Max = 7; break;
2167 case Intrinsic::mips_sat_s_h:
2168 case Intrinsic::mips_sat_u_h: Max = 15; break;
2169 case Intrinsic::mips_sat_s_w:
2170 case Intrinsic::mips_sat_u_w: Max = 31; break;
2171 case Intrinsic::mips_sat_s_d:
2172 case Intrinsic::mips_sat_u_d: Max = 63; break;
2173 default: llvm_unreachable("Unmatched intrinsic");
2174 }
2175 int64_t Value = cast<ConstantSDNode>(Op->getOperand(2))->getSExtValue();
2176 if (Value < 0 || Value > Max)
2177 report_fatal_error("Immediate out of range");
2178 return SDValue();
2179 }
Daniel Sanders26307182013-09-24 14:20:00 +00002180 case Intrinsic::mips_shf_b:
2181 case Intrinsic::mips_shf_h:
Simon Dardis548a53f2017-01-10 16:40:57 +00002182 case Intrinsic::mips_shf_w: {
2183 int64_t Value = cast<ConstantSDNode>(Op->getOperand(2))->getSExtValue();
2184 if (Value < 0 || Value > 255)
2185 report_fatal_error("Immediate out of range");
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002186 return DAG.getNode(MipsISD::SHF, DL, Op->getValueType(0),
Daniel Sanders26307182013-09-24 14:20:00 +00002187 Op->getOperand(2), Op->getOperand(1));
Simon Dardis548a53f2017-01-10 16:40:57 +00002188 }
2189 case Intrinsic::mips_sldi_b:
2190 case Intrinsic::mips_sldi_h:
2191 case Intrinsic::mips_sldi_w:
2192 case Intrinsic::mips_sldi_d: {
2193 // Report an error for out of range values.
2194 int64_t Max;
2195 switch (Intrinsic) {
2196 case Intrinsic::mips_sldi_b: Max = 15; break;
2197 case Intrinsic::mips_sldi_h: Max = 7; break;
2198 case Intrinsic::mips_sldi_w: Max = 3; break;
2199 case Intrinsic::mips_sldi_d: Max = 1; break;
2200 default: llvm_unreachable("Unmatched intrinsic");
2201 }
2202 int64_t Value = cast<ConstantSDNode>(Op->getOperand(3))->getSExtValue();
2203 if (Value < 0 || Value > Max)
2204 report_fatal_error("Immediate out of range");
2205 return SDValue();
2206 }
Daniel Sandersfbcb5822013-09-11 11:58:30 +00002207 case Intrinsic::mips_sll_b:
2208 case Intrinsic::mips_sll_h:
2209 case Intrinsic::mips_sll_w:
2210 case Intrinsic::mips_sll_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002211 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0), Op->getOperand(1),
2212 Op->getOperand(2));
Daniel Sanderscba19222013-09-24 10:28:18 +00002213 case Intrinsic::mips_slli_b:
2214 case Intrinsic::mips_slli_h:
2215 case Intrinsic::mips_slli_w:
2216 case Intrinsic::mips_slli_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002217 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0),
2218 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
Daniel Sanderse7ef0c82013-10-30 13:07:44 +00002219 case Intrinsic::mips_splat_b:
2220 case Intrinsic::mips_splat_h:
2221 case Intrinsic::mips_splat_w:
2222 case Intrinsic::mips_splat_d:
2223 // We can't lower via VECTOR_SHUFFLE because it requires constant shuffle
2224 // masks, nor can we lower via BUILD_VECTOR & EXTRACT_VECTOR_ELT because
2225 // EXTRACT_VECTOR_ELT can't extract i64's on MIPS32.
2226 // Instead we lower to MipsISD::VSHF and match from there.
2227 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
Daniel Sanders50b80412013-11-15 12:56:49 +00002228 lowerMSASplatZExt(Op, 2, DAG), Op->getOperand(1),
Daniel Sanderse7ef0c82013-10-30 13:07:44 +00002229 Op->getOperand(1));
Daniel Sanders7e51fe12013-09-27 11:48:57 +00002230 case Intrinsic::mips_splati_b:
2231 case Intrinsic::mips_splati_h:
2232 case Intrinsic::mips_splati_w:
2233 case Intrinsic::mips_splati_d:
2234 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2235 lowerMSASplatImm(Op, 2, DAG), Op->getOperand(1),
2236 Op->getOperand(1));
Daniel Sandersfbcb5822013-09-11 11:58:30 +00002237 case Intrinsic::mips_sra_b:
2238 case Intrinsic::mips_sra_h:
2239 case Intrinsic::mips_sra_w:
2240 case Intrinsic::mips_sra_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002241 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0), Op->getOperand(1),
2242 Op->getOperand(2));
Daniel Sanderscba19222013-09-24 10:28:18 +00002243 case Intrinsic::mips_srai_b:
2244 case Intrinsic::mips_srai_h:
2245 case Intrinsic::mips_srai_w:
2246 case Intrinsic::mips_srai_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002247 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0),
2248 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
Simon Dardis548a53f2017-01-10 16:40:57 +00002249 case Intrinsic::mips_srari_b:
2250 case Intrinsic::mips_srari_h:
2251 case Intrinsic::mips_srari_w:
2252 case Intrinsic::mips_srari_d: {
2253 // Report an error for out of range values.
2254 int64_t Max;
2255 switch (Intrinsic) {
2256 case Intrinsic::mips_srari_b: Max = 7; break;
2257 case Intrinsic::mips_srari_h: Max = 15; break;
2258 case Intrinsic::mips_srari_w: Max = 31; break;
2259 case Intrinsic::mips_srari_d: Max = 63; break;
2260 default: llvm_unreachable("Unmatched intrinsic");
2261 }
2262 int64_t Value = cast<ConstantSDNode>(Op->getOperand(2))->getSExtValue();
2263 if (Value < 0 || Value > Max)
2264 report_fatal_error("Immediate out of range");
2265 return SDValue();
2266 }
Daniel Sandersfbcb5822013-09-11 11:58:30 +00002267 case Intrinsic::mips_srl_b:
2268 case Intrinsic::mips_srl_h:
2269 case Intrinsic::mips_srl_w:
2270 case Intrinsic::mips_srl_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002271 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0), Op->getOperand(1),
2272 Op->getOperand(2));
Daniel Sanderscba19222013-09-24 10:28:18 +00002273 case Intrinsic::mips_srli_b:
2274 case Intrinsic::mips_srli_h:
2275 case Intrinsic::mips_srli_w:
2276 case Intrinsic::mips_srli_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002277 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0),
2278 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
Simon Dardis548a53f2017-01-10 16:40:57 +00002279 case Intrinsic::mips_srlri_b:
2280 case Intrinsic::mips_srlri_h:
2281 case Intrinsic::mips_srlri_w:
2282 case Intrinsic::mips_srlri_d: {
2283 // Report an error for out of range values.
2284 int64_t Max;
2285 switch (Intrinsic) {
2286 case Intrinsic::mips_srlri_b: Max = 7; break;
2287 case Intrinsic::mips_srlri_h: Max = 15; break;
2288 case Intrinsic::mips_srlri_w: Max = 31; break;
2289 case Intrinsic::mips_srlri_d: Max = 63; break;
2290 default: llvm_unreachable("Unmatched intrinsic");
2291 }
2292 int64_t Value = cast<ConstantSDNode>(Op->getOperand(2))->getSExtValue();
2293 if (Value < 0 || Value > Max)
2294 report_fatal_error("Immediate out of range");
2295 return SDValue();
2296 }
Daniel Sandersfbcb5822013-09-11 11:58:30 +00002297 case Intrinsic::mips_subv_b:
2298 case Intrinsic::mips_subv_h:
2299 case Intrinsic::mips_subv_w:
2300 case Intrinsic::mips_subv_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002301 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0), Op->getOperand(1),
2302 Op->getOperand(2));
Daniel Sanders86d0c8d2013-09-23 14:29:55 +00002303 case Intrinsic::mips_subvi_b:
2304 case Intrinsic::mips_subvi_h:
2305 case Intrinsic::mips_subvi_w:
2306 case Intrinsic::mips_subvi_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002307 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0),
2308 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
Daniel Sanderse5087042013-09-24 14:02:15 +00002309 case Intrinsic::mips_vshf_b:
2310 case Intrinsic::mips_vshf_h:
2311 case Intrinsic::mips_vshf_w:
2312 case Intrinsic::mips_vshf_d:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002313 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
Daniel Sanderse5087042013-09-24 14:02:15 +00002314 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
Daniel Sanders8ca81e42013-09-23 12:57:42 +00002315 case Intrinsic::mips_xor_v:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002316 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0), Op->getOperand(1),
2317 Op->getOperand(2));
Daniel Sandersbfc39ce2013-09-24 12:32:47 +00002318 case Intrinsic::mips_xori_b:
Daniel Sanders84e7caf2013-09-27 10:25:41 +00002319 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0),
2320 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
Marcin Koscielnicki7efdca52016-04-27 17:21:49 +00002321 case Intrinsic::thread_pointer: {
2322 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2323 return DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
2324 }
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00002325 }
2326}
2327
Simon Dardis548a53f2017-01-10 16:40:57 +00002328static SDValue lowerMSALoadIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr,
2329 const MipsSubtarget &Subtarget) {
Daniel Sanderse6ed5b72013-08-28 12:04:29 +00002330 SDLoc DL(Op);
2331 SDValue ChainIn = Op->getOperand(0);
2332 SDValue Address = Op->getOperand(2);
2333 SDValue Offset = Op->getOperand(3);
2334 EVT ResTy = Op->getValueType(0);
2335 EVT PtrTy = Address->getValueType(0);
2336
Simon Dardis548a53f2017-01-10 16:40:57 +00002337 // For N64 addresses have the underlying type MVT::i64. This intrinsic
2338 // however takes an i32 signed constant offset. The actual type of the
2339 // intrinsic is a scaled signed i10.
2340 if (Subtarget.isABI_N64())
2341 Offset = DAG.getNode(ISD::SIGN_EXTEND, DL, PtrTy, Offset);
2342
Daniel Sanderse6ed5b72013-08-28 12:04:29 +00002343 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
Justin Lebar9c375812016-07-15 18:27:10 +00002344 return DAG.getLoad(ResTy, DL, ChainIn, Address, MachinePointerInfo(),
2345 /* Alignment = */ 16);
Daniel Sanderse6ed5b72013-08-28 12:04:29 +00002346}
2347
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00002348SDValue MipsSETargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
2349 SelectionDAG &DAG) const {
Daniel Sanderse6ed5b72013-08-28 12:04:29 +00002350 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2351 switch (Intr) {
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00002352 default:
2353 return SDValue();
2354 case Intrinsic::mips_extp:
2355 return lowerDSPIntr(Op, DAG, MipsISD::EXTP);
2356 case Intrinsic::mips_extpdp:
2357 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP);
2358 case Intrinsic::mips_extr_w:
2359 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W);
2360 case Intrinsic::mips_extr_r_w:
2361 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W);
2362 case Intrinsic::mips_extr_rs_w:
2363 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W);
2364 case Intrinsic::mips_extr_s_h:
2365 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H);
2366 case Intrinsic::mips_mthlip:
2367 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP);
2368 case Intrinsic::mips_mulsaq_s_w_ph:
2369 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH);
2370 case Intrinsic::mips_maq_s_w_phl:
2371 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL);
2372 case Intrinsic::mips_maq_s_w_phr:
2373 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR);
2374 case Intrinsic::mips_maq_sa_w_phl:
2375 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL);
2376 case Intrinsic::mips_maq_sa_w_phr:
2377 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR);
2378 case Intrinsic::mips_dpaq_s_w_ph:
2379 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH);
2380 case Intrinsic::mips_dpsq_s_w_ph:
2381 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH);
2382 case Intrinsic::mips_dpaq_sa_l_w:
2383 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W);
2384 case Intrinsic::mips_dpsq_sa_l_w:
2385 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W);
2386 case Intrinsic::mips_dpaqx_s_w_ph:
2387 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH);
2388 case Intrinsic::mips_dpaqx_sa_w_ph:
2389 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH);
2390 case Intrinsic::mips_dpsqx_s_w_ph:
2391 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH);
2392 case Intrinsic::mips_dpsqx_sa_w_ph:
2393 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH);
Daniel Sanderse6ed5b72013-08-28 12:04:29 +00002394 case Intrinsic::mips_ld_b:
2395 case Intrinsic::mips_ld_h:
2396 case Intrinsic::mips_ld_w:
2397 case Intrinsic::mips_ld_d:
Simon Dardis548a53f2017-01-10 16:40:57 +00002398 return lowerMSALoadIntr(Op, DAG, Intr, Subtarget);
Daniel Sanderse6ed5b72013-08-28 12:04:29 +00002399 }
2400}
2401
Simon Dardis548a53f2017-01-10 16:40:57 +00002402static SDValue lowerMSAStoreIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr,
2403 const MipsSubtarget &Subtarget) {
Daniel Sanderse6ed5b72013-08-28 12:04:29 +00002404 SDLoc DL(Op);
2405 SDValue ChainIn = Op->getOperand(0);
2406 SDValue Value = Op->getOperand(2);
2407 SDValue Address = Op->getOperand(3);
2408 SDValue Offset = Op->getOperand(4);
2409 EVT PtrTy = Address->getValueType(0);
2410
Simon Dardis548a53f2017-01-10 16:40:57 +00002411 // For N64 addresses have the underlying type MVT::i64. This intrinsic
2412 // however takes an i32 signed constant offset. The actual type of the
2413 // intrinsic is a scaled signed i10.
2414 if (Subtarget.isABI_N64())
2415 Offset = DAG.getNode(ISD::SIGN_EXTEND, DL, PtrTy, Offset);
2416
Daniel Sanderse6ed5b72013-08-28 12:04:29 +00002417 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
2418
Justin Lebar9c375812016-07-15 18:27:10 +00002419 return DAG.getStore(ChainIn, DL, Value, Address, MachinePointerInfo(),
2420 /* Alignment = */ 16);
Daniel Sanderse6ed5b72013-08-28 12:04:29 +00002421}
2422
2423SDValue MipsSETargetLowering::lowerINTRINSIC_VOID(SDValue Op,
2424 SelectionDAG &DAG) const {
2425 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2426 switch (Intr) {
2427 default:
2428 return SDValue();
2429 case Intrinsic::mips_st_b:
2430 case Intrinsic::mips_st_h:
2431 case Intrinsic::mips_st_w:
2432 case Intrinsic::mips_st_d:
Simon Dardis548a53f2017-01-10 16:40:57 +00002433 return lowerMSAStoreIntr(Op, DAG, Intr, Subtarget);
Akira Hatanakaa6bbde52013-04-13 02:13:30 +00002434 }
2435}
2436
Daniel Sanders7a289d02013-09-23 12:02:46 +00002437/// \brief Check if the given BuildVectorSDNode is a splat.
2438/// This method currently relies on DAG nodes being reused when equivalent,
2439/// so it's possible for this to return false even when isConstantSplat returns
2440/// true.
2441static bool isSplatVector(const BuildVectorSDNode *N) {
Daniel Sanders7a289d02013-09-23 12:02:46 +00002442 unsigned int nOps = N->getNumOperands();
Daniel Sandersab94b532013-10-30 15:20:38 +00002443 assert(nOps > 1 && "isSplatVector has 0 or 1 sized build vector");
Daniel Sanders7a289d02013-09-23 12:02:46 +00002444
2445 SDValue Operand0 = N->getOperand(0);
2446
2447 for (unsigned int i = 1; i < nOps; ++i) {
2448 if (N->getOperand(i) != Operand0)
2449 return false;
2450 }
2451
2452 return true;
2453}
2454
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +00002455// Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT.
2456//
2457// The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We
2458// choose to sign-extend but we could have equally chosen zero-extend. The
2459// DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT
2460// result into this node later (possibly changing it to a zero-extend in the
2461// process).
2462SDValue MipsSETargetLowering::
2463lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
2464 SDLoc DL(Op);
2465 EVT ResTy = Op->getValueType(0);
2466 SDValue Op0 = Op->getOperand(0);
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00002467 EVT VecTy = Op0->getValueType(0);
2468
2469 if (!VecTy.is128BitVector())
2470 return SDValue();
2471
2472 if (ResTy.isInteger()) {
2473 SDValue Op1 = Op->getOperand(1);
2474 EVT EltTy = VecTy.getVectorElementType();
2475 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
2476 DAG.getValueType(EltTy));
2477 }
2478
2479 return Op;
Daniel Sandersa4c8f3a2013-09-23 14:03:12 +00002480}
2481
Daniel Sandersf49dd822013-09-24 13:33:07 +00002482static bool isConstantOrUndef(const SDValue Op) {
Sanjay Patel57195842016-03-14 17:28:46 +00002483 if (Op->isUndef())
Daniel Sandersf49dd822013-09-24 13:33:07 +00002484 return true;
Vasileios Kalintiris46963f62015-02-13 19:12:16 +00002485 if (isa<ConstantSDNode>(Op))
Daniel Sandersf49dd822013-09-24 13:33:07 +00002486 return true;
Vasileios Kalintiris46963f62015-02-13 19:12:16 +00002487 if (isa<ConstantFPSDNode>(Op))
Daniel Sandersf49dd822013-09-24 13:33:07 +00002488 return true;
2489 return false;
2490}
2491
2492static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) {
2493 for (unsigned i = 0; i < Op->getNumOperands(); ++i)
2494 if (isConstantOrUndef(Op->getOperand(i)))
2495 return true;
2496 return false;
2497}
2498
Daniel Sanders7a289d02013-09-23 12:02:46 +00002499// Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the
2500// backend.
2501//
2502// Lowers according to the following rules:
Daniel Sandersf49dd822013-09-24 13:33:07 +00002503// - Constant splats are legal as-is as long as the SplatBitSize is a power of
2504// 2 less than or equal to 64 and the value fits into a signed 10-bit
2505// immediate
2506// - Constant splats are lowered to bitconverted BUILD_VECTORs if SplatBitSize
2507// is a power of 2 less than or equal to 64 and the value does not fit into a
2508// signed 10-bit immediate
2509// - Non-constant splats are legal as-is.
2510// - Non-constant non-splats are lowered to sequences of INSERT_VECTOR_ELT.
2511// - All others are illegal and must be expanded.
Daniel Sanders7a289d02013-09-23 12:02:46 +00002512SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op,
2513 SelectionDAG &DAG) const {
2514 BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
2515 EVT ResTy = Op->getValueType(0);
2516 SDLoc DL(Op);
2517 APInt SplatValue, SplatUndef;
2518 unsigned SplatBitSize;
2519 bool HasAnyUndefs;
2520
Eric Christopher1c29a652014-07-18 22:55:25 +00002521 if (!Subtarget.hasMSA() || !ResTy.is128BitVector())
Daniel Sanders7a289d02013-09-23 12:02:46 +00002522 return SDValue();
2523
2524 if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
2525 HasAnyUndefs, 8,
Eric Christopher1c29a652014-07-18 22:55:25 +00002526 !Subtarget.isLittle()) && SplatBitSize <= 64) {
Daniel Sandersf49dd822013-09-24 13:33:07 +00002527 // We can only cope with 8, 16, 32, or 64-bit elements
2528 if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 &&
2529 SplatBitSize != 64)
2530 return SDValue();
2531
Simon Dardis7090d142017-03-10 13:27:14 +00002532 // If the value isn't an integer type we will have to bitcast
2533 // from an integer type first. Also, if there are any undefs, we must
2534 // lower them to defined values first.
2535 if (ResTy.isInteger() && !HasAnyUndefs)
Daniel Sandersf49dd822013-09-24 13:33:07 +00002536 return Op;
2537
2538 EVT ViaVecTy;
Daniel Sanders7a289d02013-09-23 12:02:46 +00002539
2540 switch (SplatBitSize) {
2541 default:
2542 return SDValue();
Daniel Sandersf49dd822013-09-24 13:33:07 +00002543 case 8:
2544 ViaVecTy = MVT::v16i8;
Daniel Sanders7a289d02013-09-23 12:02:46 +00002545 break;
2546 case 16:
Daniel Sandersf49dd822013-09-24 13:33:07 +00002547 ViaVecTy = MVT::v8i16;
Daniel Sanders7a289d02013-09-23 12:02:46 +00002548 break;
Daniel Sandersf49dd822013-09-24 13:33:07 +00002549 case 32:
2550 ViaVecTy = MVT::v4i32;
Daniel Sanders7a289d02013-09-23 12:02:46 +00002551 break;
Daniel Sandersf49dd822013-09-24 13:33:07 +00002552 case 64:
2553 // There's no fill.d to fall back on for 64-bit values
2554 return SDValue();
Daniel Sanders7a289d02013-09-23 12:02:46 +00002555 }
2556
Daniel Sanders50b80412013-11-15 12:56:49 +00002557 // SelectionDAG::getConstant will promote SplatValue appropriately.
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00002558 SDValue Result = DAG.getConstant(SplatValue, DL, ViaVecTy);
Daniel Sandersf49dd822013-09-24 13:33:07 +00002559
Daniel Sanders50b80412013-11-15 12:56:49 +00002560 // Bitcast to the type we originally wanted
Daniel Sandersf49dd822013-09-24 13:33:07 +00002561 if (ViaVecTy != ResTy)
2562 Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result);
Daniel Sanders7a289d02013-09-23 12:02:46 +00002563
2564 return Result;
Daniel Sandersf49dd822013-09-24 13:33:07 +00002565 } else if (isSplatVector(Node))
2566 return Op;
2567 else if (!isConstantOrUndefBUILD_VECTOR(Node)) {
Daniel Sandersf86622b2013-09-24 13:16:15 +00002568 // Use INSERT_VECTOR_ELT operations rather than expand to stores.
2569 // The resulting code is the same length as the expansion, but it doesn't
2570 // use memory operations
2571 EVT ResTy = Node->getValueType(0);
2572
2573 assert(ResTy.isVector());
2574
2575 unsigned NumElts = ResTy.getVectorNumElements();
2576 SDValue Vector = DAG.getUNDEF(ResTy);
2577 for (unsigned i = 0; i < NumElts; ++i) {
2578 Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector,
2579 Node->getOperand(i),
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00002580 DAG.getConstant(i, DL, MVT::i32));
Daniel Sandersf86622b2013-09-24 13:16:15 +00002581 }
2582 return Vector;
2583 }
Daniel Sanders7a289d02013-09-23 12:02:46 +00002584
2585 return SDValue();
2586}
2587
Daniel Sanders26307182013-09-24 14:20:00 +00002588// Lower VECTOR_SHUFFLE into SHF (if possible).
2589//
2590// SHF splits the vector into blocks of four elements, then shuffles these
2591// elements according to a <4 x i2> constant (encoded as an integer immediate).
2592//
2593// It is therefore possible to lower into SHF when the mask takes the form:
2594// <a, b, c, d, a+4, b+4, c+4, d+4, a+8, b+8, c+8, d+8, ...>
2595// When undef's appear they are treated as if they were whatever value is
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002596// necessary in order to fit the above forms.
Daniel Sanders26307182013-09-24 14:20:00 +00002597//
2598// For example:
2599// %2 = shufflevector <8 x i16> %0, <8 x i16> undef,
2600// <8 x i32> <i32 3, i32 2, i32 1, i32 0,
2601// i32 7, i32 6, i32 5, i32 4>
2602// is lowered to:
2603// (SHF_H $w0, $w1, 27)
2604// where the 27 comes from:
2605// 3 + (2 << 2) + (1 << 4) + (0 << 6)
2606static SDValue lowerVECTOR_SHUFFLE_SHF(SDValue Op, EVT ResTy,
2607 SmallVector<int, 16> Indices,
2608 SelectionDAG &DAG) {
2609 int SHFIndices[4] = { -1, -1, -1, -1 };
2610
2611 if (Indices.size() < 4)
2612 return SDValue();
2613
2614 for (unsigned i = 0; i < 4; ++i) {
2615 for (unsigned j = i; j < Indices.size(); j += 4) {
2616 int Idx = Indices[j];
2617
2618 // Convert from vector index to 4-element subvector index
2619 // If an index refers to an element outside of the subvector then give up
2620 if (Idx != -1) {
2621 Idx -= 4 * (j / 4);
2622 if (Idx < 0 || Idx >= 4)
2623 return SDValue();
2624 }
2625
2626 // If the mask has an undef, replace it with the current index.
2627 // Note that it might still be undef if the current index is also undef
2628 if (SHFIndices[i] == -1)
2629 SHFIndices[i] = Idx;
2630
2631 // Check that non-undef values are the same as in the mask. If they
2632 // aren't then give up
2633 if (!(Idx == -1 || Idx == SHFIndices[i]))
2634 return SDValue();
2635 }
2636 }
2637
2638 // Calculate the immediate. Replace any remaining undefs with zero
2639 APInt Imm(32, 0);
2640 for (int i = 3; i >= 0; --i) {
2641 int Idx = SHFIndices[i];
2642
2643 if (Idx == -1)
2644 Idx = 0;
2645
2646 Imm <<= 2;
2647 Imm |= Idx & 0x3;
2648 }
2649
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00002650 SDLoc DL(Op);
2651 return DAG.getNode(MipsISD::SHF, DL, ResTy,
2652 DAG.getConstant(Imm, DL, MVT::i32), Op->getOperand(0));
Daniel Sanders26307182013-09-24 14:20:00 +00002653}
2654
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002655/// Determine whether a range fits a regular pattern of values.
2656/// This function accounts for the possibility of jumping over the End iterator.
2657template <typename ValType>
2658static bool
2659fitsRegularPattern(typename SmallVectorImpl<ValType>::const_iterator Begin,
2660 unsigned CheckStride,
2661 typename SmallVectorImpl<ValType>::const_iterator End,
2662 ValType ExpectedIndex, unsigned ExpectedIndexStride) {
2663 auto &I = Begin;
2664
2665 while (I != End) {
2666 if (*I != -1 && *I != ExpectedIndex)
2667 return false;
2668 ExpectedIndex += ExpectedIndexStride;
2669
2670 // Incrementing past End is undefined behaviour so we must increment one
2671 // step at a time and check for End at each step.
2672 for (unsigned n = 0; n < CheckStride && I != End; ++n, ++I)
2673 ; // Empty loop body.
2674 }
2675 return true;
2676}
2677
2678// Determine whether VECTOR_SHUFFLE is a SPLATI.
2679//
2680// It is a SPLATI when the mask is:
2681// <x, x, x, ...>
2682// where x is any valid index.
2683//
2684// When undef's appear in the mask they are treated as if they were whatever
2685// value is necessary in order to fit the above form.
2686static bool isVECTOR_SHUFFLE_SPLATI(SDValue Op, EVT ResTy,
2687 SmallVector<int, 16> Indices,
2688 SelectionDAG &DAG) {
2689 assert((Indices.size() % 2) == 0);
2690
2691 int SplatIndex = -1;
2692 for (const auto &V : Indices) {
2693 if (V != -1) {
2694 SplatIndex = V;
2695 break;
2696 }
2697 }
2698
2699 return fitsRegularPattern<int>(Indices.begin(), 1, Indices.end(), SplatIndex,
2700 0);
2701}
2702
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002703// Lower VECTOR_SHUFFLE into ILVEV (if possible).
2704//
2705// ILVEV interleaves the even elements from each vector.
2706//
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002707// It is possible to lower into ILVEV when the mask consists of two of the
2708// following forms interleaved:
2709// <0, 2, 4, ...>
2710// <n, n+2, n+4, ...>
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002711// where n is the number of elements in the vector.
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002712// For example:
2713// <0, 0, 2, 2, 4, 4, ...>
2714// <0, n, 2, n+2, 4, n+4, ...>
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002715//
2716// When undef's appear in the mask they are treated as if they were whatever
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002717// value is necessary in order to fit the above forms.
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002718static SDValue lowerVECTOR_SHUFFLE_ILVEV(SDValue Op, EVT ResTy,
2719 SmallVector<int, 16> Indices,
2720 SelectionDAG &DAG) {
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002721 assert((Indices.size() % 2) == 0);
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002722
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002723 SDValue Wt;
2724 SDValue Ws;
2725 const auto &Begin = Indices.begin();
2726 const auto &End = Indices.end();
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002727
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002728 // Check even elements are taken from the even elements of one half or the
2729 // other and pick an operand accordingly.
2730 if (fitsRegularPattern<int>(Begin, 2, End, 0, 2))
2731 Wt = Op->getOperand(0);
2732 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size(), 2))
2733 Wt = Op->getOperand(1);
2734 else
2735 return SDValue();
2736
2737 // Check odd elements are taken from the even elements of one half or the
2738 // other and pick an operand accordingly.
2739 if (fitsRegularPattern<int>(Begin + 1, 2, End, 0, 2))
2740 Ws = Op->getOperand(0);
2741 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size(), 2))
2742 Ws = Op->getOperand(1);
2743 else
2744 return SDValue();
2745
2746 return DAG.getNode(MipsISD::ILVEV, SDLoc(Op), ResTy, Ws, Wt);
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002747}
2748
2749// Lower VECTOR_SHUFFLE into ILVOD (if possible).
2750//
2751// ILVOD interleaves the odd elements from each vector.
2752//
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002753// It is possible to lower into ILVOD when the mask consists of two of the
2754// following forms interleaved:
2755// <1, 3, 5, ...>
2756// <n+1, n+3, n+5, ...>
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002757// where n is the number of elements in the vector.
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002758// For example:
2759// <1, 1, 3, 3, 5, 5, ...>
2760// <1, n+1, 3, n+3, 5, n+5, ...>
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002761//
2762// When undef's appear in the mask they are treated as if they were whatever
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002763// value is necessary in order to fit the above forms.
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002764static SDValue lowerVECTOR_SHUFFLE_ILVOD(SDValue Op, EVT ResTy,
2765 SmallVector<int, 16> Indices,
2766 SelectionDAG &DAG) {
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002767 assert((Indices.size() % 2) == 0);
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002768
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002769 SDValue Wt;
2770 SDValue Ws;
2771 const auto &Begin = Indices.begin();
2772 const auto &End = Indices.end();
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002773
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002774 // Check even elements are taken from the odd elements of one half or the
2775 // other and pick an operand accordingly.
2776 if (fitsRegularPattern<int>(Begin, 2, End, 1, 2))
2777 Wt = Op->getOperand(0);
2778 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size() + 1, 2))
2779 Wt = Op->getOperand(1);
2780 else
2781 return SDValue();
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002782
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002783 // Check odd elements are taken from the odd elements of one half or the
2784 // other and pick an operand accordingly.
2785 if (fitsRegularPattern<int>(Begin + 1, 2, End, 1, 2))
2786 Ws = Op->getOperand(0);
2787 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size() + 1, 2))
2788 Ws = Op->getOperand(1);
2789 else
2790 return SDValue();
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002791
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002792 return DAG.getNode(MipsISD::ILVOD, SDLoc(Op), ResTy, Wt, Ws);
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002793}
2794
2795// Lower VECTOR_SHUFFLE into ILVR (if possible).
2796//
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002797// ILVR interleaves consecutive elements from the right (lowest-indexed) half of
2798// each vector.
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002799//
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002800// It is possible to lower into ILVR when the mask consists of two of the
2801// following forms interleaved:
2802// <0, 1, 2, ...>
2803// <n, n+1, n+2, ...>
2804// where n is the number of elements in the vector.
2805// For example:
2806// <0, 0, 1, 1, 2, 2, ...>
2807// <0, n, 1, n+1, 2, n+2, ...>
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002808//
2809// When undef's appear in the mask they are treated as if they were whatever
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002810// value is necessary in order to fit the above forms.
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002811static SDValue lowerVECTOR_SHUFFLE_ILVR(SDValue Op, EVT ResTy,
2812 SmallVector<int, 16> Indices,
2813 SelectionDAG &DAG) {
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002814 assert((Indices.size() % 2) == 0);
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002815
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002816 SDValue Wt;
2817 SDValue Ws;
2818 const auto &Begin = Indices.begin();
2819 const auto &End = Indices.end();
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002820
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002821 // Check even elements are taken from the right (lowest-indexed) elements of
2822 // one half or the other and pick an operand accordingly.
2823 if (fitsRegularPattern<int>(Begin, 2, End, 0, 1))
2824 Wt = Op->getOperand(0);
2825 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size(), 1))
2826 Wt = Op->getOperand(1);
2827 else
2828 return SDValue();
2829
2830 // Check odd elements are taken from the right (lowest-indexed) elements of
2831 // one half or the other and pick an operand accordingly.
2832 if (fitsRegularPattern<int>(Begin + 1, 2, End, 0, 1))
2833 Ws = Op->getOperand(0);
2834 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size(), 1))
2835 Ws = Op->getOperand(1);
2836 else
2837 return SDValue();
2838
2839 return DAG.getNode(MipsISD::ILVR, SDLoc(Op), ResTy, Ws, Wt);
2840}
2841
2842// Lower VECTOR_SHUFFLE into ILVL (if possible).
2843//
2844// ILVL interleaves consecutive elements from the left (highest-indexed) half
2845// of each vector.
2846//
2847// It is possible to lower into ILVL when the mask consists of two of the
2848// following forms interleaved:
2849// <x, x+1, x+2, ...>
2850// <n+x, n+x+1, n+x+2, ...>
2851// where n is the number of elements in the vector and x is half n.
2852// For example:
2853// <x, x, x+1, x+1, x+2, x+2, ...>
2854// <x, n+x, x+1, n+x+1, x+2, n+x+2, ...>
2855//
2856// When undef's appear in the mask they are treated as if they were whatever
2857// value is necessary in order to fit the above forms.
2858static SDValue lowerVECTOR_SHUFFLE_ILVL(SDValue Op, EVT ResTy,
2859 SmallVector<int, 16> Indices,
2860 SelectionDAG &DAG) {
2861 assert((Indices.size() % 2) == 0);
2862
2863 unsigned HalfSize = Indices.size() / 2;
2864 SDValue Wt;
2865 SDValue Ws;
2866 const auto &Begin = Indices.begin();
2867 const auto &End = Indices.end();
2868
2869 // Check even elements are taken from the left (highest-indexed) elements of
2870 // one half or the other and pick an operand accordingly.
2871 if (fitsRegularPattern<int>(Begin, 2, End, HalfSize, 1))
2872 Wt = Op->getOperand(0);
2873 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size() + HalfSize, 1))
2874 Wt = Op->getOperand(1);
2875 else
2876 return SDValue();
2877
2878 // Check odd elements are taken from the left (highest-indexed) elements of
2879 // one half or the other and pick an operand accordingly.
2880 if (fitsRegularPattern<int>(Begin + 1, 2, End, HalfSize, 1))
2881 Ws = Op->getOperand(0);
2882 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size() + HalfSize,
2883 1))
2884 Ws = Op->getOperand(1);
2885 else
2886 return SDValue();
2887
2888 return DAG.getNode(MipsISD::ILVL, SDLoc(Op), ResTy, Ws, Wt);
Daniel Sanders2ed228b2013-09-24 14:36:12 +00002889}
2890
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002891// Lower VECTOR_SHUFFLE into PCKEV (if possible).
2892//
2893// PCKEV copies the even elements of each vector into the result vector.
2894//
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002895// It is possible to lower into PCKEV when the mask consists of two of the
2896// following forms concatenated:
2897// <0, 2, 4, ...>
2898// <n, n+2, n+4, ...>
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002899// where n is the number of elements in the vector.
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002900// For example:
2901// <0, 2, 4, ..., 0, 2, 4, ...>
2902// <0, 2, 4, ..., n, n+2, n+4, ...>
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002903//
2904// When undef's appear in the mask they are treated as if they were whatever
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002905// value is necessary in order to fit the above forms.
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002906static SDValue lowerVECTOR_SHUFFLE_PCKEV(SDValue Op, EVT ResTy,
2907 SmallVector<int, 16> Indices,
2908 SelectionDAG &DAG) {
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002909 assert((Indices.size() % 2) == 0);
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002910
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002911 SDValue Wt;
2912 SDValue Ws;
2913 const auto &Begin = Indices.begin();
2914 const auto &Mid = Indices.begin() + Indices.size() / 2;
2915 const auto &End = Indices.end();
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002916
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002917 if (fitsRegularPattern<int>(Begin, 1, Mid, 0, 2))
2918 Wt = Op->getOperand(0);
2919 else if (fitsRegularPattern<int>(Begin, 1, Mid, Indices.size(), 2))
2920 Wt = Op->getOperand(1);
2921 else
2922 return SDValue();
2923
2924 if (fitsRegularPattern<int>(Mid, 1, End, 0, 2))
2925 Ws = Op->getOperand(0);
2926 else if (fitsRegularPattern<int>(Mid, 1, End, Indices.size(), 2))
2927 Ws = Op->getOperand(1);
2928 else
2929 return SDValue();
2930
2931 return DAG.getNode(MipsISD::PCKEV, SDLoc(Op), ResTy, Ws, Wt);
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002932}
2933
2934// Lower VECTOR_SHUFFLE into PCKOD (if possible).
2935//
2936// PCKOD copies the odd elements of each vector into the result vector.
2937//
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002938// It is possible to lower into PCKOD when the mask consists of two of the
2939// following forms concatenated:
2940// <1, 3, 5, ...>
2941// <n+1, n+3, n+5, ...>
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002942// where n is the number of elements in the vector.
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002943// For example:
2944// <1, 3, 5, ..., 1, 3, 5, ...>
2945// <1, 3, 5, ..., n+1, n+3, n+5, ...>
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002946//
2947// When undef's appear in the mask they are treated as if they were whatever
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002948// value is necessary in order to fit the above forms.
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002949static SDValue lowerVECTOR_SHUFFLE_PCKOD(SDValue Op, EVT ResTy,
2950 SmallVector<int, 16> Indices,
2951 SelectionDAG &DAG) {
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002952 assert((Indices.size() % 2) == 0);
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002953
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002954 SDValue Wt;
2955 SDValue Ws;
2956 const auto &Begin = Indices.begin();
2957 const auto &Mid = Indices.begin() + Indices.size() / 2;
2958 const auto &End = Indices.end();
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002959
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00002960 if (fitsRegularPattern<int>(Begin, 1, Mid, 1, 2))
2961 Wt = Op->getOperand(0);
2962 else if (fitsRegularPattern<int>(Begin, 1, Mid, Indices.size() + 1, 2))
2963 Wt = Op->getOperand(1);
2964 else
2965 return SDValue();
2966
2967 if (fitsRegularPattern<int>(Mid, 1, End, 1, 2))
2968 Ws = Op->getOperand(0);
2969 else if (fitsRegularPattern<int>(Mid, 1, End, Indices.size() + 1, 2))
2970 Ws = Op->getOperand(1);
2971 else
2972 return SDValue();
2973
2974 return DAG.getNode(MipsISD::PCKOD, SDLoc(Op), ResTy, Ws, Wt);
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00002975}
2976
Daniel Sanderse5087042013-09-24 14:02:15 +00002977// Lower VECTOR_SHUFFLE into VSHF.
2978//
2979// This mostly consists of converting the shuffle indices in Indices into a
2980// BUILD_VECTOR and adding it as an operand to the resulting VSHF. There is
2981// also code to eliminate unused operands of the VECTOR_SHUFFLE. For example,
2982// if the type is v8i16 and all the indices are less than 8 then the second
2983// operand is unused and can be replaced with anything. We choose to replace it
2984// with the used operand since this reduces the number of instructions overall.
2985static SDValue lowerVECTOR_SHUFFLE_VSHF(SDValue Op, EVT ResTy,
2986 SmallVector<int, 16> Indices,
2987 SelectionDAG &DAG) {
2988 SmallVector<SDValue, 16> Ops;
2989 SDValue Op0;
2990 SDValue Op1;
2991 EVT MaskVecTy = ResTy.changeVectorElementTypeToInteger();
2992 EVT MaskEltTy = MaskVecTy.getVectorElementType();
2993 bool Using1stVec = false;
2994 bool Using2ndVec = false;
2995 SDLoc DL(Op);
2996 int ResTyNumElts = ResTy.getVectorNumElements();
2997
2998 for (int i = 0; i < ResTyNumElts; ++i) {
2999 // Idx == -1 means UNDEF
3000 int Idx = Indices[i];
3001
3002 if (0 <= Idx && Idx < ResTyNumElts)
3003 Using1stVec = true;
3004 if (ResTyNumElts <= Idx && Idx < ResTyNumElts * 2)
3005 Using2ndVec = true;
3006 }
3007
3008 for (SmallVector<int, 16>::iterator I = Indices.begin(); I != Indices.end();
3009 ++I)
Sergey Dmitrouk842a51b2015-04-28 14:05:47 +00003010 Ops.push_back(DAG.getTargetConstant(*I, DL, MaskEltTy));
Daniel Sanderse5087042013-09-24 14:02:15 +00003011
Ahmed Bougacha128f8732016-04-26 21:15:30 +00003012 SDValue MaskVec = DAG.getBuildVector(MaskVecTy, DL, Ops);
Daniel Sanderse5087042013-09-24 14:02:15 +00003013
3014 if (Using1stVec && Using2ndVec) {
3015 Op0 = Op->getOperand(0);
3016 Op1 = Op->getOperand(1);
3017 } else if (Using1stVec)
3018 Op0 = Op1 = Op->getOperand(0);
3019 else if (Using2ndVec)
3020 Op0 = Op1 = Op->getOperand(1);
3021 else
3022 llvm_unreachable("shuffle vector mask references neither vector operand?");
3023
Daniel Sandersf88a29e2014-03-21 16:56:51 +00003024 // VECTOR_SHUFFLE concatenates the vectors in an vectorwise fashion.
3025 // <0b00, 0b01> + <0b10, 0b11> -> <0b00, 0b01, 0b10, 0b11>
3026 // VSHF concatenates the vectors in a bitwise fashion:
3027 // <0b00, 0b01> + <0b10, 0b11> ->
3028 // 0b0100 + 0b1110 -> 0b01001110
3029 // <0b10, 0b11, 0b00, 0b01>
3030 // We must therefore swap the operands to get the correct result.
3031 return DAG.getNode(MipsISD::VSHF, DL, ResTy, MaskVec, Op1, Op0);
Daniel Sanderse5087042013-09-24 14:02:15 +00003032}
3033
3034// Lower VECTOR_SHUFFLE into one of a number of instructions depending on the
3035// indices in the shuffle.
3036SDValue MipsSETargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
3037 SelectionDAG &DAG) const {
3038 ShuffleVectorSDNode *Node = cast<ShuffleVectorSDNode>(Op);
3039 EVT ResTy = Op->getValueType(0);
3040
3041 if (!ResTy.is128BitVector())
3042 return SDValue();
3043
3044 int ResTyNumElts = ResTy.getVectorNumElements();
3045 SmallVector<int, 16> Indices;
3046
3047 for (int i = 0; i < ResTyNumElts; ++i)
3048 Indices.push_back(Node->getMaskElt(i));
3049
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00003050 // splati.[bhwd] is preferable to the others but is matched from
3051 // MipsISD::VSHF.
3052 if (isVECTOR_SHUFFLE_SPLATI(Op, ResTy, Indices, DAG))
3053 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
Ahmed Bougachaf8dfb472016-02-09 22:54:12 +00003054 SDValue Result;
3055 if ((Result = lowerVECTOR_SHUFFLE_ILVEV(Op, ResTy, Indices, DAG)))
Daniel Sanders2ed228b2013-09-24 14:36:12 +00003056 return Result;
Ahmed Bougachaf8dfb472016-02-09 22:54:12 +00003057 if ((Result = lowerVECTOR_SHUFFLE_ILVOD(Op, ResTy, Indices, DAG)))
Daniel Sanders2ed228b2013-09-24 14:36:12 +00003058 return Result;
Ahmed Bougachaf8dfb472016-02-09 22:54:12 +00003059 if ((Result = lowerVECTOR_SHUFFLE_ILVL(Op, ResTy, Indices, DAG)))
Daniel Sanders2ed228b2013-09-24 14:36:12 +00003060 return Result;
Ahmed Bougachaf8dfb472016-02-09 22:54:12 +00003061 if ((Result = lowerVECTOR_SHUFFLE_ILVR(Op, ResTy, Indices, DAG)))
Daniel Sanders2ed228b2013-09-24 14:36:12 +00003062 return Result;
Ahmed Bougachaf8dfb472016-02-09 22:54:12 +00003063 if ((Result = lowerVECTOR_SHUFFLE_PCKEV(Op, ResTy, Indices, DAG)))
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00003064 return Result;
Ahmed Bougachaf8dfb472016-02-09 22:54:12 +00003065 if ((Result = lowerVECTOR_SHUFFLE_PCKOD(Op, ResTy, Indices, DAG)))
Daniel Sandersfae5f2a2013-09-24 14:53:25 +00003066 return Result;
Ahmed Bougachaf8dfb472016-02-09 22:54:12 +00003067 if ((Result = lowerVECTOR_SHUFFLE_SHF(Op, ResTy, Indices, DAG)))
Daniel Sandersc8cd58f2015-05-19 12:24:52 +00003068 return Result;
Daniel Sanderse5087042013-09-24 14:02:15 +00003069 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
3070}
3071
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003072MachineBasicBlock *
3073MipsSETargetLowering::emitBPOSGE32(MachineInstr &MI,
3074 MachineBasicBlock *BB) const {
Akira Hatanaka96ca1822013-03-13 00:54:29 +00003075 // $bb:
3076 // bposge32_pseudo $vr0
3077 // =>
3078 // $bb:
3079 // bposge32 $tbb
3080 // $fbb:
3081 // li $vr2, 0
3082 // b $sink
3083 // $tbb:
3084 // li $vr1, 1
3085 // $sink:
3086 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
3087
3088 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
Eric Christopher96e72c62015-01-29 23:27:36 +00003089 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Akira Hatanaka13e6ccf2013-08-06 23:08:38 +00003090 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003091 DebugLoc DL = MI.getDebugLoc();
Akira Hatanaka96ca1822013-03-13 00:54:29 +00003092 const BasicBlock *LLVM_BB = BB->getBasicBlock();
Benjamin Kramerb6d0bd42014-03-02 12:27:27 +00003093 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
Akira Hatanaka96ca1822013-03-13 00:54:29 +00003094 MachineFunction *F = BB->getParent();
3095 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
3096 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
3097 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
3098 F->insert(It, FBB);
3099 F->insert(It, TBB);
3100 F->insert(It, Sink);
3101
3102 // Transfer the remainder of BB and its successor edges to Sink.
Benjamin Kramerb6d0bd42014-03-02 12:27:27 +00003103 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
Akira Hatanaka96ca1822013-03-13 00:54:29 +00003104 BB->end());
3105 Sink->transferSuccessorsAndUpdatePHIs(BB);
3106
3107 // Add successors.
3108 BB->addSuccessor(FBB);
3109 BB->addSuccessor(TBB);
3110 FBB->addSuccessor(Sink);
3111 TBB->addSuccessor(Sink);
3112
3113 // Insert the real bposge32 instruction to $BB.
3114 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
Hrvoje Varga6f09cdf2016-05-13 11:32:53 +00003115 // Insert the real bposge32c instruction to $BB.
3116 BuildMI(BB, DL, TII->get(Mips::BPOSGE32C_MMR3)).addMBB(TBB);
Akira Hatanaka96ca1822013-03-13 00:54:29 +00003117
3118 // Fill $FBB.
3119 unsigned VR2 = RegInfo.createVirtualRegister(RC);
3120 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
3121 .addReg(Mips::ZERO).addImm(0);
3122 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
3123
3124 // Fill $TBB.
3125 unsigned VR1 = RegInfo.createVirtualRegister(RC);
3126 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
3127 .addReg(Mips::ZERO).addImm(1);
3128
3129 // Insert phi function to $Sink.
3130 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003131 MI.getOperand(0).getReg())
3132 .addReg(VR2)
3133 .addMBB(FBB)
3134 .addReg(VR1)
3135 .addMBB(TBB);
Akira Hatanaka96ca1822013-03-13 00:54:29 +00003136
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003137 MI.eraseFromParent(); // The pseudo instruction is gone now.
Akira Hatanaka96ca1822013-03-13 00:54:29 +00003138 return Sink;
3139}
Daniel Sandersce09d072013-08-28 12:14:50 +00003140
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003141MachineBasicBlock *MipsSETargetLowering::emitMSACBranchPseudo(
3142 MachineInstr &MI, MachineBasicBlock *BB, unsigned BranchOp) const {
Daniel Sandersce09d072013-08-28 12:14:50 +00003143 // $bb:
3144 // vany_nonzero $rd, $ws
3145 // =>
3146 // $bb:
3147 // bnz.b $ws, $tbb
3148 // b $fbb
3149 // $fbb:
3150 // li $rd1, 0
3151 // b $sink
3152 // $tbb:
3153 // li $rd2, 1
3154 // $sink:
3155 // $rd = phi($rd1, $fbb, $rd2, $tbb)
3156
3157 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
Eric Christopher96e72c62015-01-29 23:27:36 +00003158 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sandersce09d072013-08-28 12:14:50 +00003159 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003160 DebugLoc DL = MI.getDebugLoc();
Daniel Sandersce09d072013-08-28 12:14:50 +00003161 const BasicBlock *LLVM_BB = BB->getBasicBlock();
Benjamin Kramerb6d0bd42014-03-02 12:27:27 +00003162 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
Daniel Sandersce09d072013-08-28 12:14:50 +00003163 MachineFunction *F = BB->getParent();
3164 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
3165 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
3166 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
3167 F->insert(It, FBB);
3168 F->insert(It, TBB);
3169 F->insert(It, Sink);
3170
3171 // Transfer the remainder of BB and its successor edges to Sink.
Benjamin Kramerb6d0bd42014-03-02 12:27:27 +00003172 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
Daniel Sandersce09d072013-08-28 12:14:50 +00003173 BB->end());
3174 Sink->transferSuccessorsAndUpdatePHIs(BB);
3175
3176 // Add successors.
3177 BB->addSuccessor(FBB);
3178 BB->addSuccessor(TBB);
3179 FBB->addSuccessor(Sink);
3180 TBB->addSuccessor(Sink);
3181
3182 // Insert the real bnz.b instruction to $BB.
3183 BuildMI(BB, DL, TII->get(BranchOp))
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003184 .addReg(MI.getOperand(1).getReg())
3185 .addMBB(TBB);
Daniel Sandersce09d072013-08-28 12:14:50 +00003186
3187 // Fill $FBB.
3188 unsigned RD1 = RegInfo.createVirtualRegister(RC);
3189 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), RD1)
3190 .addReg(Mips::ZERO).addImm(0);
3191 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
3192
3193 // Fill $TBB.
3194 unsigned RD2 = RegInfo.createVirtualRegister(RC);
3195 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), RD2)
3196 .addReg(Mips::ZERO).addImm(1);
3197
3198 // Insert phi function to $Sink.
3199 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003200 MI.getOperand(0).getReg())
3201 .addReg(RD1)
3202 .addMBB(FBB)
3203 .addReg(RD2)
3204 .addMBB(TBB);
Daniel Sandersce09d072013-08-28 12:14:50 +00003205
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003206 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sandersce09d072013-08-28 12:14:50 +00003207 return Sink;
3208}
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003209
3210// Emit the COPY_FW pseudo instruction.
3211//
3212// copy_fw_pseudo $fd, $ws, n
3213// =>
3214// copy_u_w $rt, $ws, $n
3215// mtc1 $rt, $fd
3216//
3217// When n is zero, the equivalent operation can be performed with (potentially)
3218// zero instructions due to register overlaps. This optimization is never valid
3219// for lane 1 because it would require FR=0 mode which isn't supported by MSA.
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003220MachineBasicBlock *
3221MipsSETargetLowering::emitCOPY_FW(MachineInstr &MI,
3222 MachineBasicBlock *BB) const {
Eric Christopher96e72c62015-01-29 23:27:36 +00003223 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003224 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003225 DebugLoc DL = MI.getDebugLoc();
3226 unsigned Fd = MI.getOperand(0).getReg();
3227 unsigned Ws = MI.getOperand(1).getReg();
3228 unsigned Lane = MI.getOperand(2).getImm();
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003229
Daniel Sandersafe27c72015-02-23 17:22:16 +00003230 if (Lane == 0) {
3231 unsigned Wt = Ws;
3232 if (!Subtarget.useOddSPReg()) {
3233 // We must copy to an even-numbered MSA register so that the
3234 // single-precision sub-register is also guaranteed to be even-numbered.
3235 Wt = RegInfo.createVirtualRegister(&Mips::MSA128WEvensRegClass);
3236
3237 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Wt).addReg(Ws);
3238 }
3239
3240 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
3241 } else {
3242 unsigned Wt = RegInfo.createVirtualRegister(
3243 Subtarget.useOddSPReg() ? &Mips::MSA128WRegClass :
3244 &Mips::MSA128WEvensRegClass);
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003245
Daniel Sandersd9207702014-03-04 13:54:30 +00003246 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wt).addReg(Ws).addImm(Lane);
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003247 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
3248 }
3249
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003250 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003251 return BB;
3252}
3253
3254// Emit the COPY_FD pseudo instruction.
3255//
3256// copy_fd_pseudo $fd, $ws, n
3257// =>
3258// splati.d $wt, $ws, $n
3259// copy $fd, $wt:sub_64
3260//
3261// When n is zero, the equivalent operation can be performed with (potentially)
3262// zero instructions due to register overlaps. This optimization is always
3263// valid because FR=1 mode which is the only supported mode in MSA.
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003264MachineBasicBlock *
3265MipsSETargetLowering::emitCOPY_FD(MachineInstr &MI,
3266 MachineBasicBlock *BB) const {
Eric Christopher1c29a652014-07-18 22:55:25 +00003267 assert(Subtarget.isFP64bit());
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003268
Eric Christopher96e72c62015-01-29 23:27:36 +00003269 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003270 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003271 unsigned Fd = MI.getOperand(0).getReg();
3272 unsigned Ws = MI.getOperand(1).getReg();
3273 unsigned Lane = MI.getOperand(2).getImm() * 2;
3274 DebugLoc DL = MI.getDebugLoc();
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003275
3276 if (Lane == 0)
3277 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_64);
3278 else {
3279 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3280
3281 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wt).addReg(Ws).addImm(1);
3282 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_64);
3283 }
3284
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003285 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sanders39bb8ba2013-09-27 12:17:32 +00003286 return BB;
3287}
Daniel Sandersa5150702013-09-27 12:31:32 +00003288
3289// Emit the INSERT_FW pseudo instruction.
3290//
3291// insert_fw_pseudo $wd, $wd_in, $n, $fs
3292// =>
3293// subreg_to_reg $wt:sub_lo, $fs
3294// insve_w $wd[$n], $wd_in, $wt[0]
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003295MachineBasicBlock *
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003296MipsSETargetLowering::emitINSERT_FW(MachineInstr &MI,
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003297 MachineBasicBlock *BB) const {
Eric Christopher96e72c62015-01-29 23:27:36 +00003298 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sandersa5150702013-09-27 12:31:32 +00003299 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003300 DebugLoc DL = MI.getDebugLoc();
3301 unsigned Wd = MI.getOperand(0).getReg();
3302 unsigned Wd_in = MI.getOperand(1).getReg();
3303 unsigned Lane = MI.getOperand(2).getImm();
3304 unsigned Fs = MI.getOperand(3).getReg();
Daniel Sandersafe27c72015-02-23 17:22:16 +00003305 unsigned Wt = RegInfo.createVirtualRegister(
3306 Subtarget.useOddSPReg() ? &Mips::MSA128WRegClass :
3307 &Mips::MSA128WEvensRegClass);
Daniel Sandersa5150702013-09-27 12:31:32 +00003308
3309 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003310 .addImm(0)
3311 .addReg(Fs)
3312 .addImm(Mips::sub_lo);
Daniel Sandersa5150702013-09-27 12:31:32 +00003313 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_W), Wd)
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003314 .addReg(Wd_in)
3315 .addImm(Lane)
Daniel Sandersb50ccf82014-04-01 10:35:28 +00003316 .addReg(Wt)
3317 .addImm(0);
Daniel Sandersa5150702013-09-27 12:31:32 +00003318
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003319 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sandersa5150702013-09-27 12:31:32 +00003320 return BB;
3321}
3322
3323// Emit the INSERT_FD pseudo instruction.
3324//
3325// insert_fd_pseudo $wd, $fs, n
3326// =>
3327// subreg_to_reg $wt:sub_64, $fs
3328// insve_d $wd[$n], $wd_in, $wt[0]
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003329MachineBasicBlock *
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003330MipsSETargetLowering::emitINSERT_FD(MachineInstr &MI,
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003331 MachineBasicBlock *BB) const {
Eric Christopher1c29a652014-07-18 22:55:25 +00003332 assert(Subtarget.isFP64bit());
Daniel Sandersa5150702013-09-27 12:31:32 +00003333
Eric Christopher96e72c62015-01-29 23:27:36 +00003334 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sandersa5150702013-09-27 12:31:32 +00003335 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003336 DebugLoc DL = MI.getDebugLoc();
3337 unsigned Wd = MI.getOperand(0).getReg();
3338 unsigned Wd_in = MI.getOperand(1).getReg();
3339 unsigned Lane = MI.getOperand(2).getImm();
3340 unsigned Fs = MI.getOperand(3).getReg();
Daniel Sandersa5150702013-09-27 12:31:32 +00003341 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3342
3343 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003344 .addImm(0)
3345 .addReg(Fs)
3346 .addImm(Mips::sub_64);
Daniel Sandersa5150702013-09-27 12:31:32 +00003347 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_D), Wd)
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003348 .addReg(Wd_in)
3349 .addImm(Lane)
Daniel Sandersb50ccf82014-04-01 10:35:28 +00003350 .addReg(Wt)
3351 .addImm(0);
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003352
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003353 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003354 return BB;
3355}
3356
Daniel Sanderse296a0f2014-04-30 12:09:32 +00003357// Emit the INSERT_([BHWD]|F[WD])_VIDX pseudo instruction.
3358//
3359// For integer:
3360// (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $rs)
3361// =>
3362// (SLL $lanetmp1, $lane, <log2size)
3363// (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3364// (INSERT_[BHWD], $wdtmp2, $wdtmp1, 0, $rs)
3365// (NEG $lanetmp2, $lanetmp1)
3366// (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
3367//
3368// For floating point:
3369// (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $fs)
3370// =>
3371// (SUBREG_TO_REG $wt, $fs, <subreg>)
3372// (SLL $lanetmp1, $lane, <log2size)
3373// (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3374// (INSVE_[WD], $wdtmp2, 0, $wdtmp1, 0)
3375// (NEG $lanetmp2, $lanetmp1)
3376// (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003377MachineBasicBlock *MipsSETargetLowering::emitINSERT_DF_VIDX(
3378 MachineInstr &MI, MachineBasicBlock *BB, unsigned EltSizeInBytes,
3379 bool IsFP) const {
Eric Christopher96e72c62015-01-29 23:27:36 +00003380 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sanderse296a0f2014-04-30 12:09:32 +00003381 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003382 DebugLoc DL = MI.getDebugLoc();
3383 unsigned Wd = MI.getOperand(0).getReg();
3384 unsigned SrcVecReg = MI.getOperand(1).getReg();
3385 unsigned LaneReg = MI.getOperand(2).getReg();
3386 unsigned SrcValReg = MI.getOperand(3).getReg();
Daniel Sanderse296a0f2014-04-30 12:09:32 +00003387
3388 const TargetRegisterClass *VecRC = nullptr;
Daniel Sandersd3bb2082016-06-15 08:43:23 +00003389 // FIXME: This should be true for N32 too.
Eric Christopherbf33a3c2014-07-02 23:18:40 +00003390 const TargetRegisterClass *GPRRC =
Daniel Sanders4160c802015-05-05 08:48:35 +00003391 Subtarget.isABI_N64() ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
Daniel Sandersd3bb2082016-06-15 08:43:23 +00003392 unsigned SubRegIdx = Subtarget.isABI_N64() ? Mips::sub_32 : 0;
3393 unsigned ShiftOp = Subtarget.isABI_N64() ? Mips::DSLL : Mips::SLL;
Daniel Sanderse296a0f2014-04-30 12:09:32 +00003394 unsigned EltLog2Size;
3395 unsigned InsertOp = 0;
3396 unsigned InsveOp = 0;
3397 switch (EltSizeInBytes) {
3398 default:
3399 llvm_unreachable("Unexpected size");
3400 case 1:
3401 EltLog2Size = 0;
3402 InsertOp = Mips::INSERT_B;
3403 InsveOp = Mips::INSVE_B;
3404 VecRC = &Mips::MSA128BRegClass;
3405 break;
3406 case 2:
3407 EltLog2Size = 1;
3408 InsertOp = Mips::INSERT_H;
3409 InsveOp = Mips::INSVE_H;
3410 VecRC = &Mips::MSA128HRegClass;
3411 break;
3412 case 4:
3413 EltLog2Size = 2;
3414 InsertOp = Mips::INSERT_W;
3415 InsveOp = Mips::INSVE_W;
3416 VecRC = &Mips::MSA128WRegClass;
3417 break;
3418 case 8:
3419 EltLog2Size = 3;
3420 InsertOp = Mips::INSERT_D;
3421 InsveOp = Mips::INSVE_D;
3422 VecRC = &Mips::MSA128DRegClass;
3423 break;
3424 }
3425
3426 if (IsFP) {
3427 unsigned Wt = RegInfo.createVirtualRegister(VecRC);
3428 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
3429 .addImm(0)
3430 .addReg(SrcValReg)
3431 .addImm(EltSizeInBytes == 8 ? Mips::sub_64 : Mips::sub_lo);
3432 SrcValReg = Wt;
3433 }
3434
3435 // Convert the lane index into a byte index
3436 if (EltSizeInBytes != 1) {
3437 unsigned LaneTmp1 = RegInfo.createVirtualRegister(GPRRC);
Daniel Sandersd3bb2082016-06-15 08:43:23 +00003438 BuildMI(*BB, MI, DL, TII->get(ShiftOp), LaneTmp1)
Daniel Sanderse296a0f2014-04-30 12:09:32 +00003439 .addReg(LaneReg)
3440 .addImm(EltLog2Size);
3441 LaneReg = LaneTmp1;
3442 }
3443
3444 // Rotate bytes around so that the desired lane is element zero
3445 unsigned WdTmp1 = RegInfo.createVirtualRegister(VecRC);
3446 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), WdTmp1)
3447 .addReg(SrcVecReg)
3448 .addReg(SrcVecReg)
Daniel Sandersd3bb2082016-06-15 08:43:23 +00003449 .addReg(LaneReg, 0, SubRegIdx);
Daniel Sanderse296a0f2014-04-30 12:09:32 +00003450
3451 unsigned WdTmp2 = RegInfo.createVirtualRegister(VecRC);
3452 if (IsFP) {
3453 // Use insve.df to insert to element zero
3454 BuildMI(*BB, MI, DL, TII->get(InsveOp), WdTmp2)
3455 .addReg(WdTmp1)
3456 .addImm(0)
3457 .addReg(SrcValReg)
3458 .addImm(0);
3459 } else {
3460 // Use insert.df to insert to element zero
3461 BuildMI(*BB, MI, DL, TII->get(InsertOp), WdTmp2)
3462 .addReg(WdTmp1)
3463 .addReg(SrcValReg)
3464 .addImm(0);
3465 }
3466
3467 // Rotate elements the rest of the way for a full rotation.
3468 // sld.df inteprets $rt modulo the number of columns so we only need to negate
3469 // the lane index to do this.
3470 unsigned LaneTmp2 = RegInfo.createVirtualRegister(GPRRC);
Daniel Sanders4160c802015-05-05 08:48:35 +00003471 BuildMI(*BB, MI, DL, TII->get(Subtarget.isABI_N64() ? Mips::DSUB : Mips::SUB),
3472 LaneTmp2)
3473 .addReg(Subtarget.isABI_N64() ? Mips::ZERO_64 : Mips::ZERO)
Daniel Sanderse296a0f2014-04-30 12:09:32 +00003474 .addReg(LaneReg);
3475 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), Wd)
3476 .addReg(WdTmp2)
3477 .addReg(WdTmp2)
Daniel Sandersd3bb2082016-06-15 08:43:23 +00003478 .addReg(LaneTmp2, 0, SubRegIdx);
Daniel Sanderse296a0f2014-04-30 12:09:32 +00003479
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003480 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sanderse296a0f2014-04-30 12:09:32 +00003481 return BB;
3482}
3483
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003484// Emit the FILL_FW pseudo instruction.
3485//
3486// fill_fw_pseudo $wd, $fs
3487// =>
3488// implicit_def $wt1
3489// insert_subreg $wt2:subreg_lo, $wt1, $fs
3490// splati.w $wd, $wt2[0]
3491MachineBasicBlock *
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003492MipsSETargetLowering::emitFILL_FW(MachineInstr &MI,
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003493 MachineBasicBlock *BB) const {
Eric Christopher96e72c62015-01-29 23:27:36 +00003494 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003495 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003496 DebugLoc DL = MI.getDebugLoc();
3497 unsigned Wd = MI.getOperand(0).getReg();
3498 unsigned Fs = MI.getOperand(1).getReg();
Simon Dardis0e9e2372017-01-10 15:53:10 +00003499 unsigned Wt1 = RegInfo.createVirtualRegister(
3500 Subtarget.useOddSPReg() ? &Mips::MSA128WRegClass
3501 : &Mips::MSA128WEvensRegClass);
3502 unsigned Wt2 = RegInfo.createVirtualRegister(
3503 Subtarget.useOddSPReg() ? &Mips::MSA128WRegClass
3504 : &Mips::MSA128WEvensRegClass);
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003505
3506 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3507 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3508 .addReg(Wt1)
3509 .addReg(Fs)
3510 .addImm(Mips::sub_lo);
3511 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wd).addReg(Wt2).addImm(0);
3512
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003513 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003514 return BB;
3515}
3516
3517// Emit the FILL_FD pseudo instruction.
3518//
3519// fill_fd_pseudo $wd, $fs
3520// =>
3521// implicit_def $wt1
3522// insert_subreg $wt2:subreg_64, $wt1, $fs
3523// splati.d $wd, $wt2[0]
3524MachineBasicBlock *
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003525MipsSETargetLowering::emitFILL_FD(MachineInstr &MI,
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003526 MachineBasicBlock *BB) const {
Eric Christopher1c29a652014-07-18 22:55:25 +00003527 assert(Subtarget.isFP64bit());
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003528
Eric Christopher96e72c62015-01-29 23:27:36 +00003529 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003530 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003531 DebugLoc DL = MI.getDebugLoc();
3532 unsigned Wd = MI.getOperand(0).getReg();
3533 unsigned Fs = MI.getOperand(1).getReg();
Daniel Sanders1dfddc72013-10-15 13:14:41 +00003534 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3535 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3536
3537 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3538 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3539 .addReg(Wt1)
3540 .addReg(Fs)
3541 .addImm(Mips::sub_64);
3542 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wd).addReg(Wt2).addImm(0);
Daniel Sandersa5150702013-09-27 12:31:32 +00003543
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003544 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sandersa5150702013-09-27 12:31:32 +00003545 return BB;
3546}
Daniel Sandersa9521602013-10-23 10:36:52 +00003547
Simon Dardis0e2ee3b2016-11-18 16:17:44 +00003548// Emit the ST_F16_PSEDUO instruction to store a f16 value from an MSA
3549// register.
3550//
3551// STF16 MSA128F16:$wd, mem_simm10:$addr
3552// =>
3553// copy_u.h $rtemp,$wd[0]
3554// sh $rtemp, $addr
3555//
3556// Safety: We can't use st.h & co as they would over write the memory after
3557// the destination. It would require half floats be allocated 16 bytes(!) of
3558// space.
3559MachineBasicBlock *
3560MipsSETargetLowering::emitST_F16_PSEUDO(MachineInstr &MI,
3561 MachineBasicBlock *BB) const {
3562
3563 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3564 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3565 DebugLoc DL = MI.getDebugLoc();
3566 unsigned Ws = MI.getOperand(0).getReg();
3567 unsigned Rt = MI.getOperand(1).getReg();
3568 const MachineMemOperand &MMO = **MI.memoperands_begin();
3569 unsigned Imm = MMO.getOffset();
3570
3571 // Caution: A load via the GOT can expand to a GPR32 operand, a load via
3572 // spill and reload can expand as a GPR64 operand. Examine the
3573 // operand in detail and default to ABI.
3574 const TargetRegisterClass *RC =
3575 MI.getOperand(1).isReg() ? RegInfo.getRegClass(MI.getOperand(1).getReg())
3576 : (Subtarget.isABI_O32() ? &Mips::GPR32RegClass
3577 : &Mips::GPR64RegClass);
3578 const bool UsingMips32 = RC == &Mips::GPR32RegClass;
3579 unsigned Rs = RegInfo.createVirtualRegister(RC);
3580
3581 BuildMI(*BB, MI, DL, TII->get(Mips::COPY_U_H), Rs).addReg(Ws).addImm(0);
3582 BuildMI(*BB, MI, DL, TII->get(UsingMips32 ? Mips::SH : Mips::SH64))
3583 .addReg(Rs)
3584 .addReg(Rt)
3585 .addImm(Imm)
3586 .addMemOperand(BB->getParent()->getMachineMemOperand(
3587 &MMO, MMO.getOffset(), MMO.getSize()));
3588
3589 MI.eraseFromParent();
3590 return BB;
3591}
3592
3593// Emit the LD_F16_PSEDUO instruction to load a f16 value into an MSA register.
3594//
3595// LD_F16 MSA128F16:$wd, mem_simm10:$addr
3596// =>
3597// lh $rtemp, $addr
3598// fill.h $wd, $rtemp
3599//
3600// Safety: We can't use ld.h & co as they over-read from the source.
3601// Additionally, if the address is not modulo 16, 2 cases can occur:
3602// a) Segmentation fault as the load instruction reads from a memory page
3603// memory it's not supposed to.
3604// b) The load crosses an implementation specific boundary, requiring OS
3605// intervention.
3606//
3607MachineBasicBlock *
3608MipsSETargetLowering::emitLD_F16_PSEUDO(MachineInstr &MI,
3609 MachineBasicBlock *BB) const {
3610
3611 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3612 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3613 DebugLoc DL = MI.getDebugLoc();
3614 unsigned Wd = MI.getOperand(0).getReg();
3615
3616 // Caution: A load via the GOT can expand to a GPR32 operand, a load via
3617 // spill and reload can expand as a GPR64 operand. Examine the
3618 // operand in detail and default to ABI.
3619 const TargetRegisterClass *RC =
3620 MI.getOperand(1).isReg() ? RegInfo.getRegClass(MI.getOperand(1).getReg())
3621 : (Subtarget.isABI_O32() ? &Mips::GPR32RegClass
3622 : &Mips::GPR64RegClass);
3623
3624 const bool UsingMips32 = RC == &Mips::GPR32RegClass;
3625 unsigned Rt = RegInfo.createVirtualRegister(RC);
3626
3627 MachineInstrBuilder MIB =
3628 BuildMI(*BB, MI, DL, TII->get(UsingMips32 ? Mips::LH : Mips::LH64), Rt);
3629 for (unsigned i = 1; i < MI.getNumOperands(); i++)
Diana Picus116bbab2017-01-13 09:58:52 +00003630 MIB.add(MI.getOperand(i));
Simon Dardis0e2ee3b2016-11-18 16:17:44 +00003631
3632 BuildMI(*BB, MI, DL, TII->get(Mips::FILL_H), Wd).addReg(Rt);
3633
3634 MI.eraseFromParent();
3635 return BB;
3636}
3637
3638// Emit the FPROUND_PSEUDO instruction.
3639//
3640// Round an FGR64Opnd, FGR32Opnd to an f16.
3641//
3642// Safety: Cycle the operand through the GPRs so the result always ends up
3643// the correct MSA register.
3644//
3645// FIXME: This copying is strictly unnecessary. If we could tie FGR32Opnd:$Fs
3646// / FGR64Opnd:$Fs and MSA128F16:$Wd to the same physical register
3647// (which they can be, as the MSA registers are defined to alias the
3648// FPU's 64 bit and 32 bit registers) the result can be accessed using
3649// the correct register class. That requires operands be tie-able across
3650// register classes which have a sub/super register class relationship.
3651//
3652// For FPG32Opnd:
3653//
3654// FPROUND MSA128F16:$wd, FGR32Opnd:$fs
3655// =>
3656// mfc1 $rtemp, $fs
3657// fill.w $rtemp, $wtemp
3658// fexdo.w $wd, $wtemp, $wtemp
3659//
3660// For FPG64Opnd on mips32r2+:
3661//
3662// FPROUND MSA128F16:$wd, FGR64Opnd:$fs
3663// =>
3664// mfc1 $rtemp, $fs
3665// fill.w $rtemp, $wtemp
3666// mfhc1 $rtemp2, $fs
3667// insert.w $wtemp[1], $rtemp2
3668// insert.w $wtemp[3], $rtemp2
3669// fexdo.w $wtemp2, $wtemp, $wtemp
3670// fexdo.h $wd, $temp2, $temp2
3671//
3672// For FGR64Opnd on mips64r2+:
3673//
3674// FPROUND MSA128F16:$wd, FGR64Opnd:$fs
3675// =>
3676// dmfc1 $rtemp, $fs
3677// fill.d $rtemp, $wtemp
3678// fexdo.w $wtemp2, $wtemp, $wtemp
3679// fexdo.h $wd, $wtemp2, $wtemp2
3680//
3681// Safety note: As $wtemp is UNDEF, we may provoke a spurious exception if the
3682// undef bits are "just right" and the exception enable bits are
3683// set. By using fill.w to replicate $fs into all elements over
3684// insert.w for one element, we avoid that potiential case. If
3685// fexdo.[hw] causes an exception in, the exception is valid and it
3686// occurs for all elements.
3687//
3688MachineBasicBlock *
3689MipsSETargetLowering::emitFPROUND_PSEUDO(MachineInstr &MI,
3690 MachineBasicBlock *BB,
3691 bool IsFGR64) const {
3692
3693 // Strictly speaking, we need MIPS32R5 to support MSA. We'll be generous
3694 // here. It's technically doable to support MIPS32 here, but the ISA forbids
3695 // it.
3696 assert(Subtarget.hasMSA() && Subtarget.hasMips32r2());
3697
3698 bool IsFGR64onMips64 = Subtarget.hasMips64() && IsFGR64;
3699
3700 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3701 DebugLoc DL = MI.getDebugLoc();
3702 unsigned Wd = MI.getOperand(0).getReg();
3703 unsigned Fs = MI.getOperand(1).getReg();
3704
3705 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3706 unsigned Wtemp = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3707 const TargetRegisterClass *GPRRC =
3708 IsFGR64onMips64 ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
3709 unsigned MFC1Opc = IsFGR64onMips64 ? Mips::DMFC1 : Mips::MFC1;
3710 unsigned FILLOpc = IsFGR64onMips64 ? Mips::FILL_D : Mips::FILL_W;
3711
3712 // Perform the register class copy as mentioned above.
3713 unsigned Rtemp = RegInfo.createVirtualRegister(GPRRC);
3714 BuildMI(*BB, MI, DL, TII->get(MFC1Opc), Rtemp).addReg(Fs);
3715 BuildMI(*BB, MI, DL, TII->get(FILLOpc), Wtemp).addReg(Rtemp);
3716 unsigned WPHI = Wtemp;
3717
3718 if (!Subtarget.hasMips64() && IsFGR64) {
3719 unsigned Rtemp2 = RegInfo.createVirtualRegister(GPRRC);
3720 BuildMI(*BB, MI, DL, TII->get(Mips::MFHC1_D64), Rtemp2).addReg(Fs);
3721 unsigned Wtemp2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3722 unsigned Wtemp3 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3723 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_W), Wtemp2)
3724 .addReg(Wtemp)
3725 .addReg(Rtemp2)
3726 .addImm(1);
3727 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_W), Wtemp3)
3728 .addReg(Wtemp2)
3729 .addReg(Rtemp2)
3730 .addImm(3);
3731 WPHI = Wtemp3;
3732 }
3733
3734 if (IsFGR64) {
3735 unsigned Wtemp2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3736 BuildMI(*BB, MI, DL, TII->get(Mips::FEXDO_W), Wtemp2)
3737 .addReg(WPHI)
3738 .addReg(WPHI);
3739 WPHI = Wtemp2;
3740 }
3741
3742 BuildMI(*BB, MI, DL, TII->get(Mips::FEXDO_H), Wd).addReg(WPHI).addReg(WPHI);
3743
3744 MI.eraseFromParent();
3745 return BB;
3746}
3747
3748// Emit the FPEXTEND_PSEUDO instruction.
3749//
3750// Expand an f16 to either a FGR32Opnd or FGR64Opnd.
3751//
3752// Safety: Cycle the result through the GPRs so the result always ends up
3753// the correct floating point register.
3754//
3755// FIXME: This copying is strictly unnecessary. If we could tie FGR32Opnd:$Fd
3756// / FGR64Opnd:$Fd and MSA128F16:$Ws to the same physical register
3757// (which they can be, as the MSA registers are defined to alias the
3758// FPU's 64 bit and 32 bit registers) the result can be accessed using
3759// the correct register class. That requires operands be tie-able across
3760// register classes which have a sub/super register class relationship. I
3761// haven't checked.
3762//
3763// For FGR32Opnd:
3764//
3765// FPEXTEND FGR32Opnd:$fd, MSA128F16:$ws
3766// =>
3767// fexupr.w $wtemp, $ws
3768// copy_s.w $rtemp, $ws[0]
3769// mtc1 $rtemp, $fd
3770//
3771// For FGR64Opnd on Mips64:
3772//
3773// FPEXTEND FGR64Opnd:$fd, MSA128F16:$ws
3774// =>
3775// fexupr.w $wtemp, $ws
3776// fexupr.d $wtemp2, $wtemp
3777// copy_s.d $rtemp, $wtemp2s[0]
3778// dmtc1 $rtemp, $fd
3779//
3780// For FGR64Opnd on Mips32:
3781//
3782// FPEXTEND FGR64Opnd:$fd, MSA128F16:$ws
3783// =>
3784// fexupr.w $wtemp, $ws
3785// fexupr.d $wtemp2, $wtemp
3786// copy_s.w $rtemp, $wtemp2[0]
3787// mtc1 $rtemp, $ftemp
3788// copy_s.w $rtemp2, $wtemp2[1]
3789// $fd = mthc1 $rtemp2, $ftemp
3790//
3791MachineBasicBlock *
3792MipsSETargetLowering::emitFPEXTEND_PSEUDO(MachineInstr &MI,
3793 MachineBasicBlock *BB,
3794 bool IsFGR64) const {
3795
3796 // Strictly speaking, we need MIPS32R5 to support MSA. We'll be generous
3797 // here. It's technically doable to support MIPS32 here, but the ISA forbids
3798 // it.
3799 assert(Subtarget.hasMSA() && Subtarget.hasMips32r2());
3800
3801 bool IsFGR64onMips64 = Subtarget.hasMips64() && IsFGR64;
3802 bool IsFGR64onMips32 = !Subtarget.hasMips64() && IsFGR64;
3803
3804 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3805 DebugLoc DL = MI.getDebugLoc();
3806 unsigned Fd = MI.getOperand(0).getReg();
3807 unsigned Ws = MI.getOperand(1).getReg();
3808
3809 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3810 const TargetRegisterClass *GPRRC =
3811 IsFGR64onMips64 ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
3812 unsigned MTC1Opc = IsFGR64onMips64 ? Mips::DMTC1 : Mips::MTC1;
3813 unsigned COPYOpc = IsFGR64onMips64 ? Mips::COPY_S_D : Mips::COPY_S_W;
3814
3815 unsigned Wtemp = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3816 unsigned WPHI = Wtemp;
3817
3818 BuildMI(*BB, MI, DL, TII->get(Mips::FEXUPR_W), Wtemp).addReg(Ws);
3819 if (IsFGR64) {
3820 WPHI = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3821 BuildMI(*BB, MI, DL, TII->get(Mips::FEXUPR_D), WPHI).addReg(Wtemp);
3822 }
3823
3824 // Perform the safety regclass copy mentioned above.
3825 unsigned Rtemp = RegInfo.createVirtualRegister(GPRRC);
3826 unsigned FPRPHI = IsFGR64onMips32
3827 ? RegInfo.createVirtualRegister(&Mips::FGR64RegClass)
3828 : Fd;
3829 BuildMI(*BB, MI, DL, TII->get(COPYOpc), Rtemp).addReg(WPHI).addImm(0);
3830 BuildMI(*BB, MI, DL, TII->get(MTC1Opc), FPRPHI).addReg(Rtemp);
3831
3832 if (IsFGR64onMips32) {
3833 unsigned Rtemp2 = RegInfo.createVirtualRegister(GPRRC);
3834 BuildMI(*BB, MI, DL, TII->get(Mips::COPY_S_W), Rtemp2)
3835 .addReg(WPHI)
3836 .addImm(1);
3837 BuildMI(*BB, MI, DL, TII->get(Mips::MTHC1_D64), Fd)
3838 .addReg(FPRPHI)
3839 .addReg(Rtemp2);
3840 }
3841
3842 MI.eraseFromParent();
3843 return BB;
3844}
3845
Daniel Sandersa9521602013-10-23 10:36:52 +00003846// Emit the FEXP2_W_1 pseudo instructions.
3847//
3848// fexp2_w_1_pseudo $wd, $wt
3849// =>
3850// ldi.w $ws, 1
3851// fexp2.w $wd, $ws, $wt
3852MachineBasicBlock *
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003853MipsSETargetLowering::emitFEXP2_W_1(MachineInstr &MI,
Daniel Sandersa9521602013-10-23 10:36:52 +00003854 MachineBasicBlock *BB) const {
Eric Christopher96e72c62015-01-29 23:27:36 +00003855 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sandersa9521602013-10-23 10:36:52 +00003856 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3857 const TargetRegisterClass *RC = &Mips::MSA128WRegClass;
3858 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3859 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003860 DebugLoc DL = MI.getDebugLoc();
Daniel Sandersa9521602013-10-23 10:36:52 +00003861
3862 // Splat 1.0 into a vector
3863 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_W), Ws1).addImm(1);
3864 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_W), Ws2).addReg(Ws1);
3865
3866 // Emit 1.0 * fexp2(Wt)
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003867 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_W), MI.getOperand(0).getReg())
Daniel Sandersa9521602013-10-23 10:36:52 +00003868 .addReg(Ws2)
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003869 .addReg(MI.getOperand(1).getReg());
Daniel Sandersa9521602013-10-23 10:36:52 +00003870
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003871 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sandersa9521602013-10-23 10:36:52 +00003872 return BB;
3873}
3874
3875// Emit the FEXP2_D_1 pseudo instructions.
3876//
3877// fexp2_d_1_pseudo $wd, $wt
3878// =>
3879// ldi.d $ws, 1
3880// fexp2.d $wd, $ws, $wt
3881MachineBasicBlock *
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003882MipsSETargetLowering::emitFEXP2_D_1(MachineInstr &MI,
Daniel Sandersa9521602013-10-23 10:36:52 +00003883 MachineBasicBlock *BB) const {
Eric Christopher96e72c62015-01-29 23:27:36 +00003884 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
Daniel Sandersa9521602013-10-23 10:36:52 +00003885 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3886 const TargetRegisterClass *RC = &Mips::MSA128DRegClass;
3887 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3888 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003889 DebugLoc DL = MI.getDebugLoc();
Daniel Sandersa9521602013-10-23 10:36:52 +00003890
3891 // Splat 1.0 into a vector
3892 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_D), Ws1).addImm(1);
3893 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_D), Ws2).addReg(Ws1);
3894
3895 // Emit 1.0 * fexp2(Wt)
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003896 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_D), MI.getOperand(0).getReg())
Daniel Sandersa9521602013-10-23 10:36:52 +00003897 .addReg(Ws2)
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003898 .addReg(MI.getOperand(1).getReg());
Daniel Sandersa9521602013-10-23 10:36:52 +00003899
Duncan P. N. Exon Smithe4f5e4f2016-06-30 22:52:52 +00003900 MI.eraseFromParent(); // The pseudo instruction is gone now.
Daniel Sandersa9521602013-10-23 10:36:52 +00003901 return BB;
3902}