Jia Liu | e1d6196 | 2012-02-19 02:03:36 +0000 | [diff] [blame] | 1 | //===-- X86InstrFragmentsSIMD.td - x86 SIMD ISA ------------*- tablegen -*-===// |
David Greene | 509be1f | 2010-02-09 23:52:19 +0000 | [diff] [blame] | 2 | // |
| 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. |
Jia Liu | b22310f | 2012-02-18 12:03:15 +0000 | [diff] [blame] | 7 | // |
David Greene | 509be1f | 2010-02-09 23:52:19 +0000 | [diff] [blame] | 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // This file provides pattern fragments useful for SIMD instructions. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | //===----------------------------------------------------------------------===// |
Bruno Cardoso Lopes | e446aef | 2015-02-05 13:22:50 +0000 | [diff] [blame] | 15 | // MMX specific DAG Nodes. |
| 16 | //===----------------------------------------------------------------------===// |
| 17 | |
| 18 | // Low word of MMX to GPR. |
| 19 | def MMX_X86movd2w : SDNode<"X86ISD::MMX_MOVD2W", SDTypeProfile<1, 1, |
| 20 | [SDTCisVT<0, i32>, SDTCisVT<1, x86mmx>]>>; |
Bruno Cardoso Lopes | ab9ae87 | 2015-02-05 13:23:07 +0000 | [diff] [blame] | 21 | // GPR to low word of MMX. |
| 22 | def MMX_X86movw2d : SDNode<"X86ISD::MMX_MOVW2D", SDTypeProfile<1, 1, |
| 23 | [SDTCisVT<0, x86mmx>, SDTCisVT<1, i32>]>>; |
Bruno Cardoso Lopes | e446aef | 2015-02-05 13:22:50 +0000 | [diff] [blame] | 24 | |
| 25 | //===----------------------------------------------------------------------===// |
David Greene | 509be1f | 2010-02-09 23:52:19 +0000 | [diff] [blame] | 26 | // MMX Pattern Fragments |
| 27 | //===----------------------------------------------------------------------===// |
| 28 | |
Dale Johannesen | dd224d2 | 2010-09-30 23:57:10 +0000 | [diff] [blame] | 29 | def load_mmx : PatFrag<(ops node:$ptr), (x86mmx (load node:$ptr))>; |
| 30 | def bc_mmx : PatFrag<(ops node:$in), (x86mmx (bitconvert node:$in))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 31 | |
| 32 | //===----------------------------------------------------------------------===// |
| 33 | // SSE specific DAG Nodes. |
| 34 | //===----------------------------------------------------------------------===// |
| 35 | |
| 36 | def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>, |
| 37 | SDTCisFP<0>, SDTCisInt<2> ]>; |
| 38 | def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>, |
Craig Topper | aefaab6 | 2014-01-26 04:59:39 +0000 | [diff] [blame] | 39 | SDTCisFP<1>, SDTCisVT<3, i8>, |
| 40 | SDTCisVec<1>]>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 41 | |
Benjamin Kramer | 4669d18 | 2012-12-21 14:04:55 +0000 | [diff] [blame] | 42 | def X86umin : SDNode<"X86ISD::UMIN", SDTIntBinOp>; |
| 43 | def X86umax : SDNode<"X86ISD::UMAX", SDTIntBinOp>; |
| 44 | def X86smin : SDNode<"X86ISD::SMIN", SDTIntBinOp>; |
| 45 | def X86smax : SDNode<"X86ISD::SMAX", SDTIntBinOp>; |
| 46 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 47 | def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>; |
| 48 | def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>; |
Nadav Rotem | 178250a | 2012-08-19 13:06:16 +0000 | [diff] [blame] | 49 | |
| 50 | // Commutative and Associative FMIN and FMAX. |
| 51 | def X86fminc : SDNode<"X86ISD::FMINC", SDTFPBinOp, |
| 52 | [SDNPCommutative, SDNPAssociative]>; |
| 53 | def X86fmaxc : SDNode<"X86ISD::FMAXC", SDTFPBinOp, |
| 54 | [SDNPCommutative, SDNPAssociative]>; |
| 55 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 56 | def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp, |
| 57 | [SDNPCommutative, SDNPAssociative]>; |
| 58 | def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp, |
| 59 | [SDNPCommutative, SDNPAssociative]>; |
| 60 | def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp, |
| 61 | [SDNPCommutative, SDNPAssociative]>; |
Benjamin Kramer | 5bc180c | 2013-08-04 12:05:16 +0000 | [diff] [blame] | 62 | def X86fandn : SDNode<"X86ISD::FANDN", SDTFPBinOp, |
| 63 | [SDNPCommutative, SDNPAssociative]>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 64 | def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>; |
| 65 | def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>; |
| 66 | def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>; |
Stuart Hastings | 9f20804 | 2011-06-01 04:39:42 +0000 | [diff] [blame] | 67 | def X86fgetsign: SDNode<"X86ISD::FGETSIGNx86",SDTFPToIntOp>; |
Duncan Sands | 0e4fcb8 | 2011-09-22 20:15:48 +0000 | [diff] [blame] | 68 | def X86fhadd : SDNode<"X86ISD::FHADD", SDTFPBinOp>; |
| 69 | def X86fhsub : SDNode<"X86ISD::FHSUB", SDTFPBinOp>; |
Craig Topper | f984efb | 2011-11-19 09:02:40 +0000 | [diff] [blame] | 70 | def X86hadd : SDNode<"X86ISD::HADD", SDTIntBinOp>; |
| 71 | def X86hsub : SDNode<"X86ISD::HSUB", SDTIntBinOp>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 72 | def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>; |
| 73 | def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>; |
Elena Demikhovsky | 47fc44e | 2013-12-16 13:52:35 +0000 | [diff] [blame] | 74 | def X86cmps : SDNode<"X86ISD::FSETCC", SDTX86Cmps>; |
| 75 | //def X86cmpsd : SDNode<"X86ISD::FSETCCsd", SDTX86Cmpsd>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 76 | def X86pshufb : SDNode<"X86ISD::PSHUFB", |
Craig Topper | 7834900 | 2012-01-25 06:43:11 +0000 | [diff] [blame] | 77 | SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 78 | SDTCisSameAs<0,2>]>>; |
Bruno Cardoso Lopes | 7ba479d | 2011-07-13 21:36:47 +0000 | [diff] [blame] | 79 | def X86andnp : SDNode<"X86ISD::ANDNP", |
Bruno Cardoso Lopes | 9613b64 | 2011-07-13 21:36:51 +0000 | [diff] [blame] | 80 | SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
Nate Begeman | 97b72c9 | 2010-12-17 22:55:37 +0000 | [diff] [blame] | 81 | SDTCisSameAs<0,2>]>>; |
Craig Topper | 81390be | 2011-11-19 07:33:10 +0000 | [diff] [blame] | 82 | def X86psign : SDNode<"X86ISD::PSIGN", |
Craig Topper | de6b73b | 2011-11-19 07:07:26 +0000 | [diff] [blame] | 83 | SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
Nate Begeman | 97b72c9 | 2010-12-17 22:55:37 +0000 | [diff] [blame] | 84 | SDTCisSameAs<0,2>]>>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 85 | def X86pextrb : SDNode<"X86ISD::PEXTRB", |
| 86 | SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>; |
| 87 | def X86pextrw : SDNode<"X86ISD::PEXTRW", |
| 88 | SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>; |
| 89 | def X86pinsrb : SDNode<"X86ISD::PINSRB", |
| 90 | SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>, |
| 91 | SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>; |
| 92 | def X86pinsrw : SDNode<"X86ISD::PINSRW", |
| 93 | SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>, |
| 94 | SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>; |
Filipe Cabecinhas | 2035221 | 2014-04-21 20:07:29 +0000 | [diff] [blame] | 95 | def X86insertps : SDNode<"X86ISD::INSERTPS", |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 96 | SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>, |
Chandler Carruth | 373b2b1 | 2014-09-06 10:00:01 +0000 | [diff] [blame] | 97 | SDTCisVT<2, v4f32>, SDTCisVT<3, i8>]>>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 98 | def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL", |
| 99 | SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>; |
Elena Demikhovsky | 8d7e56c | 2012-04-22 09:39:03 +0000 | [diff] [blame] | 100 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 101 | def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad, |
Chris Lattner | 54e5329 | 2010-09-22 00:34:38 +0000 | [diff] [blame] | 102 | [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>; |
Michael Liao | 34107b9 | 2012-08-14 21:24:47 +0000 | [diff] [blame] | 103 | |
Michael Liao | 1be96bb | 2012-10-23 17:34:00 +0000 | [diff] [blame] | 104 | def X86vzext : SDNode<"X86ISD::VZEXT", |
| 105 | SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>, |
Craig Topper | aefaab6 | 2014-01-26 04:59:39 +0000 | [diff] [blame] | 106 | SDTCisInt<0>, SDTCisInt<1>, |
| 107 | SDTCisOpSmallerThanOp<1, 0>]>>; |
Michael Liao | 1be96bb | 2012-10-23 17:34:00 +0000 | [diff] [blame] | 108 | |
| 109 | def X86vsext : SDNode<"X86ISD::VSEXT", |
| 110 | SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>, |
Craig Topper | aefaab6 | 2014-01-26 04:59:39 +0000 | [diff] [blame] | 111 | SDTCisInt<0>, SDTCisInt<1>, |
| 112 | SDTCisOpSmallerThanOp<1, 0>]>>; |
Michael Liao | 1be96bb | 2012-10-23 17:34:00 +0000 | [diff] [blame] | 113 | |
Elena Demikhovsky | 980c6b0 | 2013-08-29 11:56:53 +0000 | [diff] [blame] | 114 | def X86vtrunc : SDNode<"X86ISD::VTRUNC", |
| 115 | SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>, |
Craig Topper | aefaab6 | 2014-01-26 04:59:39 +0000 | [diff] [blame] | 116 | SDTCisInt<0>, SDTCisInt<1>, |
| 117 | SDTCisOpSmallerThanOp<0, 1>]>>; |
Elena Demikhovsky | c5f6726 | 2013-12-17 08:33:15 +0000 | [diff] [blame] | 118 | def X86trunc : SDNode<"X86ISD::TRUNC", |
Craig Topper | aefaab6 | 2014-01-26 04:59:39 +0000 | [diff] [blame] | 119 | SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisInt<1>, |
| 120 | SDTCisOpSmallerThanOp<0, 1>]>>; |
| 121 | |
Elena Demikhovsky | 980c6b0 | 2013-08-29 11:56:53 +0000 | [diff] [blame] | 122 | def X86vtruncm : SDNode<"X86ISD::VTRUNCM", |
| 123 | SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>, |
| 124 | SDTCisInt<0>, SDTCisInt<1>, |
Craig Topper | aefaab6 | 2014-01-26 04:59:39 +0000 | [diff] [blame] | 125 | SDTCisVec<2>, SDTCisInt<2>, |
| 126 | SDTCisOpSmallerThanOp<0, 2>]>>; |
Michael Liao | 34107b9 | 2012-08-14 21:24:47 +0000 | [diff] [blame] | 127 | def X86vfpext : SDNode<"X86ISD::VFPEXT", |
| 128 | SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>, |
Craig Topper | aefaab6 | 2014-01-26 04:59:39 +0000 | [diff] [blame] | 129 | SDTCisFP<0>, SDTCisFP<1>, |
| 130 | SDTCisOpSmallerThanOp<1, 0>]>>; |
Michael Liao | e999b86 | 2012-10-10 16:53:28 +0000 | [diff] [blame] | 131 | def X86vfpround: SDNode<"X86ISD::VFPROUND", |
| 132 | SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>, |
Craig Topper | aefaab6 | 2014-01-26 04:59:39 +0000 | [diff] [blame] | 133 | SDTCisFP<0>, SDTCisFP<1>, |
| 134 | SDTCisOpSmallerThanOp<0, 1>]>>; |
Michael Liao | 34107b9 | 2012-08-14 21:24:47 +0000 | [diff] [blame] | 135 | |
Craig Topper | 0946264 | 2012-01-22 19:15:14 +0000 | [diff] [blame] | 136 | def X86vshldq : SDNode<"X86ISD::VSHLDQ", SDTIntShiftOp>; |
| 137 | def X86vshrdq : SDNode<"X86ISD::VSRLDQ", SDTIntShiftOp>; |
Craig Topper | 0b7ad76 | 2012-01-22 23:36:02 +0000 | [diff] [blame] | 138 | def X86cmpp : SDNode<"X86ISD::CMPP", SDTX86VFCMP>; |
Craig Topper | bd488437 | 2012-01-22 22:42:16 +0000 | [diff] [blame] | 139 | def X86pcmpeq : SDNode<"X86ISD::PCMPEQ", SDTIntBinOp, [SDNPCommutative]>; |
| 140 | def X86pcmpgt : SDNode<"X86ISD::PCMPGT", SDTIntBinOp>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 141 | |
Elena Demikhovsky | 60b1f28 | 2013-08-13 13:24:07 +0000 | [diff] [blame] | 142 | def X86IntCmpMask : SDTypeProfile<1, 2, |
| 143 | [SDTCisVec<0>, SDTCisSameAs<1, 2>, SDTCisInt<1>]>; |
| 144 | def X86pcmpeqm : SDNode<"X86ISD::PCMPEQM", X86IntCmpMask, [SDNPCommutative]>; |
| 145 | def X86pcmpgtm : SDNode<"X86ISD::PCMPGTM", X86IntCmpMask>; |
| 146 | |
Elena Demikhovsky | 47fc44e | 2013-12-16 13:52:35 +0000 | [diff] [blame] | 147 | def X86CmpMaskCC : |
Craig Topper | aefaab6 | 2014-01-26 04:59:39 +0000 | [diff] [blame] | 148 | SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisInt<0>, SDTCisVec<1>, |
| 149 | SDTCisSameAs<1, 2>, SDTCisVT<3, i8>]>; |
Elena Demikhovsky | 47fc44e | 2013-12-16 13:52:35 +0000 | [diff] [blame] | 150 | def X86CmpMaskCCScalar : |
| 151 | SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>, SDTCisVT<3, i8>]>; |
| 152 | |
Elena Demikhovsky | 60b1f28 | 2013-08-13 13:24:07 +0000 | [diff] [blame] | 153 | def X86cmpm : SDNode<"X86ISD::CMPM", X86CmpMaskCC>; |
| 154 | def X86cmpmu : SDNode<"X86ISD::CMPMU", X86CmpMaskCC>; |
Elena Demikhovsky | 47fc44e | 2013-12-16 13:52:35 +0000 | [diff] [blame] | 155 | def X86cmpms : SDNode<"X86ISD::FSETCC", X86CmpMaskCCScalar>; |
Elena Demikhovsky | 60b1f28 | 2013-08-13 13:24:07 +0000 | [diff] [blame] | 156 | |
Craig Topper | 0946264 | 2012-01-22 19:15:14 +0000 | [diff] [blame] | 157 | def X86vshl : SDNode<"X86ISD::VSHL", |
| 158 | SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
| 159 | SDTCisVec<2>]>>; |
| 160 | def X86vsrl : SDNode<"X86ISD::VSRL", |
| 161 | SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
| 162 | SDTCisVec<2>]>>; |
| 163 | def X86vsra : SDNode<"X86ISD::VSRA", |
| 164 | SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
| 165 | SDTCisVec<2>]>>; |
| 166 | |
| 167 | def X86vshli : SDNode<"X86ISD::VSHLI", SDTIntShiftOp>; |
| 168 | def X86vsrli : SDNode<"X86ISD::VSRLI", SDTIntShiftOp>; |
| 169 | def X86vsrai : SDNode<"X86ISD::VSRAI", SDTIntShiftOp>; |
| 170 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 171 | def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, |
Bruno Cardoso Lopes | 91d61df | 2010-08-10 23:25:42 +0000 | [diff] [blame] | 172 | SDTCisVec<1>, |
| 173 | SDTCisSameAs<2, 1>]>; |
Benjamin Kramer | b16ccde | 2012-12-15 16:47:44 +0000 | [diff] [blame] | 174 | def X86subus : SDNode<"X86ISD::SUBUS", SDTIntBinOp>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 175 | def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>; |
Bruno Cardoso Lopes | 91d61df | 2010-08-10 23:25:42 +0000 | [diff] [blame] | 176 | def X86testp : SDNode<"X86ISD::TESTP", SDTX86CmpPTest>; |
Elena Demikhovsky | 40864b6 | 2013-08-05 08:52:21 +0000 | [diff] [blame] | 177 | def X86kortest : SDNode<"X86ISD::KORTEST", SDTX86CmpPTest>; |
Elena Demikhovsky | 47fc44e | 2013-12-16 13:52:35 +0000 | [diff] [blame] | 178 | def X86testm : SDNode<"X86ISD::TESTM", SDTypeProfile<1, 2, [SDTCisVec<0>, |
Elena Demikhovsky | 33d447a | 2013-08-21 09:36:02 +0000 | [diff] [blame] | 179 | SDTCisVec<1>, |
| 180 | SDTCisSameAs<2, 1>]>>; |
Elena Demikhovsky | a30e437 | 2014-02-05 07:05:03 +0000 | [diff] [blame] | 181 | def X86testnm : SDNode<"X86ISD::TESTNM", SDTypeProfile<1, 2, [SDTCisVec<0>, |
| 182 | SDTCisVec<1>, |
| 183 | SDTCisSameAs<2, 1>]>>; |
Elena Demikhovsky | 47fc44e | 2013-12-16 13:52:35 +0000 | [diff] [blame] | 184 | def X86select : SDNode<"X86ISD::SELECT" , SDTSelect>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 185 | |
Craig Topper | 1d471e3 | 2012-02-05 03:14:49 +0000 | [diff] [blame] | 186 | def X86pmuludq : SDNode<"X86ISD::PMULUDQ", |
| 187 | SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>, |
| 188 | SDTCisSameAs<1,2>]>>; |
Benjamin Kramer | 6d2dff6 | 2014-04-26 14:12:19 +0000 | [diff] [blame] | 189 | def X86pmuldq : SDNode<"X86ISD::PMULDQ", |
| 190 | SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>, |
| 191 | SDTCisSameAs<1,2>]>>; |
Craig Topper | 1d471e3 | 2012-02-05 03:14:49 +0000 | [diff] [blame] | 192 | |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 193 | // Specific shuffle nodes - At some point ISD::VECTOR_SHUFFLE will always get |
| 194 | // translated into one of the target nodes below during lowering. |
| 195 | // Note: this is a work in progress... |
| 196 | def SDTShuff1Op : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0,1>]>; |
| 197 | def SDTShuff2Op : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
| 198 | SDTCisSameAs<0,2>]>; |
Elena Demikhovsky | cf5b145 | 2013-08-11 07:55:09 +0000 | [diff] [blame] | 199 | def SDTShuff3Op : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
| 200 | SDTCisSameAs<0,2>, SDTCisSameAs<0,3>]>; |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 201 | |
Chandler Carruth | 6d5916a | 2014-09-23 10:08:29 +0000 | [diff] [blame] | 202 | def SDTShuff2OpM : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
| 203 | SDTCisVec<2>]>; |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 204 | def SDTShuff2OpI : SDTypeProfile<1, 2, [SDTCisVec<0>, |
| 205 | SDTCisSameAs<0,1>, SDTCisInt<2>]>; |
| 206 | def SDTShuff3OpI : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
| 207 | SDTCisSameAs<0,2>, SDTCisInt<3>]>; |
| 208 | |
Elena Demikhovsky | 45c54ad | 2013-08-07 12:34:55 +0000 | [diff] [blame] | 209 | def SDTVBroadcast : SDTypeProfile<1, 1, [SDTCisVec<0>]>; |
| 210 | def SDTVBroadcastm : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>]>; |
| 211 | |
Nadav Rotem | 9bc178a | 2012-04-11 06:40:27 +0000 | [diff] [blame] | 212 | def SDTBlend : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>, |
Chandler Carruth | 373b2b1 | 2014-09-06 10:00:01 +0000 | [diff] [blame] | 213 | SDTCisSameAs<1,2>, SDTCisVT<3, i8>]>; |
Elena Demikhovsky | 3cb3b00 | 2012-08-01 12:06:00 +0000 | [diff] [blame] | 214 | |
Elena Demikhovsky | 714f23b | 2015-02-18 07:59:20 +0000 | [diff] [blame^] | 215 | def SDTFPBinOpRound : SDTypeProfile<1, 3, [ // fadd_round, fmul_round, etc. |
| 216 | SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisFP<0>, SDTCisInt<3>]>; |
| 217 | |
Elena Demikhovsky | 3cb3b00 | 2012-08-01 12:06:00 +0000 | [diff] [blame] | 218 | def SDTFma : SDTypeProfile<1, 3, [SDTCisSameAs<0,1>, |
| 219 | SDTCisSameAs<1,2>, SDTCisSameAs<1,3>]>; |
Elena Demikhovsky | 7b0dd39 | 2015-01-28 10:21:27 +0000 | [diff] [blame] | 220 | def SDTFmaRound : SDTypeProfile<1, 4, [SDTCisSameAs<0,1>, |
| 221 | SDTCisSameAs<1,2>, SDTCisSameAs<1,3>, SDTCisInt<4>]>; |
Elena Demikhovsky | be8808d | 2014-11-12 07:31:03 +0000 | [diff] [blame] | 222 | def STDFp1SrcRm : SDTypeProfile<1, 2, [SDTCisSameAs<0,1>, |
| 223 | SDTCisVec<0>, SDTCisInt<2>]>; |
Elena Demikhovsky | 905a5a6 | 2014-11-26 10:46:49 +0000 | [diff] [blame] | 224 | def STDFp2SrcRm : SDTypeProfile<1, 3, [SDTCisSameAs<0,1>, |
| 225 | SDTCisVec<0>, SDTCisInt<3>]>; |
Bruno Cardoso Lopes | be5e987 | 2011-08-17 02:29:19 +0000 | [diff] [blame] | 226 | |
Craig Topper | 8fb09f0 | 2013-01-28 06:48:25 +0000 | [diff] [blame] | 227 | def X86PAlignr : SDNode<"X86ISD::PALIGNR", SDTShuff3OpI>; |
Adam Nemet | 2f10cc6 | 2014-08-05 17:22:55 +0000 | [diff] [blame] | 228 | def X86VAlign : SDNode<"X86ISD::VALIGN", SDTShuff3OpI>; |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 229 | |
| 230 | def X86PShufd : SDNode<"X86ISD::PSHUFD", SDTShuff2OpI>; |
| 231 | def X86PShufhw : SDNode<"X86ISD::PSHUFHW", SDTShuff2OpI>; |
| 232 | def X86PShuflw : SDNode<"X86ISD::PSHUFLW", SDTShuff2OpI>; |
| 233 | |
Craig Topper | 6e54ba7 | 2011-12-31 23:50:21 +0000 | [diff] [blame] | 234 | def X86Shufp : SDNode<"X86ISD::SHUFP", SDTShuff3OpI>; |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 235 | |
| 236 | def X86Movddup : SDNode<"X86ISD::MOVDDUP", SDTShuff1Op>; |
| 237 | def X86Movshdup : SDNode<"X86ISD::MOVSHDUP", SDTShuff1Op>; |
| 238 | def X86Movsldup : SDNode<"X86ISD::MOVSLDUP", SDTShuff1Op>; |
| 239 | |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 240 | def X86Movsd : SDNode<"X86ISD::MOVSD", SDTShuff2Op>; |
| 241 | def X86Movss : SDNode<"X86ISD::MOVSS", SDTShuff2Op>; |
| 242 | |
| 243 | def X86Movlhps : SDNode<"X86ISD::MOVLHPS", SDTShuff2Op>; |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 244 | def X86Movlhpd : SDNode<"X86ISD::MOVLHPD", SDTShuff2Op>; |
Bruno Cardoso Lopes | 03e4c35 | 2010-08-31 21:15:21 +0000 | [diff] [blame] | 245 | def X86Movhlps : SDNode<"X86ISD::MOVHLPS", SDTShuff2Op>; |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 246 | |
Bruno Cardoso Lopes | b382521 | 2010-09-01 05:08:25 +0000 | [diff] [blame] | 247 | def X86Movlps : SDNode<"X86ISD::MOVLPS", SDTShuff2Op>; |
| 248 | def X86Movlpd : SDNode<"X86ISD::MOVLPD", SDTShuff2Op>; |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 249 | |
Chandler Carruth | 8366ceb | 2014-06-20 01:05:28 +0000 | [diff] [blame] | 250 | def SDTPack : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<2, 1>]>; |
| 251 | def X86Packss : SDNode<"X86ISD::PACKSS", SDTPack>; |
| 252 | def X86Packus : SDNode<"X86ISD::PACKUS", SDTPack>; |
| 253 | |
Craig Topper | 8d4ba19 | 2011-12-06 08:21:25 +0000 | [diff] [blame] | 254 | def X86Unpckl : SDNode<"X86ISD::UNPCKL", SDTShuff2Op>; |
| 255 | def X86Unpckh : SDNode<"X86ISD::UNPCKH", SDTShuff2Op>; |
Bruno Cardoso Lopes | 6f3b38a | 2010-08-20 22:55:05 +0000 | [diff] [blame] | 256 | |
Chandler Carruth | 6d5916a | 2014-09-23 10:08:29 +0000 | [diff] [blame] | 257 | def X86VPermilpv : SDNode<"X86ISD::VPERMILPV", SDTShuff2OpM>; |
Chandler Carruth | ed5dfff | 2014-09-22 22:29:42 +0000 | [diff] [blame] | 258 | def X86VPermilpi : SDNode<"X86ISD::VPERMILPI", SDTShuff2OpI>; |
| 259 | def X86VPermv : SDNode<"X86ISD::VPERMV", SDTShuff2Op>; |
| 260 | def X86VPermi : SDNode<"X86ISD::VPERMI", SDTShuff2OpI>; |
| 261 | def X86VPermv3 : SDNode<"X86ISD::VPERMV3", SDTShuff3Op>; |
Elena Demikhovsky | a5d38a3 | 2014-01-23 14:27:26 +0000 | [diff] [blame] | 262 | def X86VPermiv3 : SDNode<"X86ISD::VPERMIV3", SDTShuff3Op>; |
Bruno Cardoso Lopes | b878caa | 2011-07-21 01:55:47 +0000 | [diff] [blame] | 263 | |
Craig Topper | 0a672ea | 2011-11-30 07:47:51 +0000 | [diff] [blame] | 264 | def X86VPerm2x128 : SDNode<"X86ISD::VPERM2X128", SDTShuff3OpI>; |
Bruno Cardoso Lopes | f15dfe5 | 2011-08-12 21:48:26 +0000 | [diff] [blame] | 265 | |
Bruno Cardoso Lopes | be5e987 | 2011-08-17 02:29:19 +0000 | [diff] [blame] | 266 | def X86VBroadcast : SDNode<"X86ISD::VBROADCAST", SDTVBroadcast>; |
Elena Demikhovsky | 45c54ad | 2013-08-07 12:34:55 +0000 | [diff] [blame] | 267 | def X86VBroadcastm : SDNode<"X86ISD::VBROADCASTM", SDTVBroadcastm>; |
Elena Demikhovsky | 8952974 | 2013-09-12 08:55:00 +0000 | [diff] [blame] | 268 | def X86Vinsert : SDNode<"X86ISD::VINSERT", SDTypeProfile<1, 3, |
| 269 | [SDTCisSameAs<0, 1>, SDTCisPtrTy<3>]>, []>; |
Elena Demikhovsky | 9f423d6 | 2014-02-10 07:02:39 +0000 | [diff] [blame] | 270 | def X86Vextract : SDNode<"X86ISD::VEXTRACT", SDTypeProfile<1, 2, |
| 271 | [SDTCisVec<1>, SDTCisPtrTy<2>]>, []>; |
Bruno Cardoso Lopes | be5e987 | 2011-08-17 02:29:19 +0000 | [diff] [blame] | 272 | |
Elena Demikhovsky | cd3c1c4 | 2012-12-05 09:24:57 +0000 | [diff] [blame] | 273 | def X86Blendi : SDNode<"X86ISD::BLENDI", SDTBlend>; |
Chandler Carruth | 204ad4c | 2014-09-15 20:09:47 +0000 | [diff] [blame] | 274 | |
| 275 | def X86Addsub : SDNode<"X86ISD::ADDSUB", SDTFPBinOp>; |
| 276 | |
Elena Demikhovsky | 714f23b | 2015-02-18 07:59:20 +0000 | [diff] [blame^] | 277 | def X86faddRnd : SDNode<"X86ISD::FADD_RND", SDTFPBinOpRound>; |
| 278 | def X86fsubRnd : SDNode<"X86ISD::FSUB_RND", SDTFPBinOpRound>; |
| 279 | def X86fmulRnd : SDNode<"X86ISD::FMUL_RND", SDTFPBinOpRound>; |
| 280 | def X86fdivRnd : SDNode<"X86ISD::FDIV_RND", SDTFPBinOpRound>; |
| 281 | |
Elena Demikhovsky | 3cb3b00 | 2012-08-01 12:06:00 +0000 | [diff] [blame] | 282 | def X86Fmadd : SDNode<"X86ISD::FMADD", SDTFma>; |
| 283 | def X86Fnmadd : SDNode<"X86ISD::FNMADD", SDTFma>; |
| 284 | def X86Fmsub : SDNode<"X86ISD::FMSUB", SDTFma>; |
| 285 | def X86Fnmsub : SDNode<"X86ISD::FNMSUB", SDTFma>; |
Craig Topper | a999c66 | 2012-08-29 07:18:25 +0000 | [diff] [blame] | 286 | def X86Fmaddsub : SDNode<"X86ISD::FMADDSUB", SDTFma>; |
| 287 | def X86Fmsubadd : SDNode<"X86ISD::FMSUBADD", SDTFma>; |
Nadav Rotem | 9bc178a | 2012-04-11 06:40:27 +0000 | [diff] [blame] | 288 | |
Elena Demikhovsky | 7b0dd39 | 2015-01-28 10:21:27 +0000 | [diff] [blame] | 289 | def X86FmaddRnd : SDNode<"X86ISD::FMADD_RND", SDTFmaRound>; |
| 290 | def X86FnmaddRnd : SDNode<"X86ISD::FNMADD_RND", SDTFmaRound>; |
| 291 | def X86FmsubRnd : SDNode<"X86ISD::FMSUB_RND", SDTFmaRound>; |
| 292 | def X86FnmsubRnd : SDNode<"X86ISD::FNMSUB_RND", SDTFmaRound>; |
| 293 | def X86FmaddsubRnd : SDNode<"X86ISD::FMADDSUB_RND", SDTFmaRound>; |
| 294 | def X86FmsubaddRnd : SDNode<"X86ISD::FMSUBADD_RND", SDTFmaRound>; |
| 295 | |
Elena Demikhovsky | be8808d | 2014-11-12 07:31:03 +0000 | [diff] [blame] | 296 | def X86rsqrt28 : SDNode<"X86ISD::RSQRT28", STDFp1SrcRm>; |
| 297 | def X86rcp28 : SDNode<"X86ISD::RCP28", STDFp1SrcRm>; |
Elena Demikhovsky | 905a5a6 | 2014-11-26 10:46:49 +0000 | [diff] [blame] | 298 | def X86exp2 : SDNode<"X86ISD::EXP2", STDFp1SrcRm>; |
| 299 | |
| 300 | def X86rsqrt28s : SDNode<"X86ISD::RSQRT28", STDFp2SrcRm>; |
| 301 | def X86rcp28s : SDNode<"X86ISD::RCP28", STDFp2SrcRm>; |
Elena Demikhovsky | be8808d | 2014-11-12 07:31:03 +0000 | [diff] [blame] | 302 | |
Craig Topper | ab47fe4 | 2012-08-06 06:22:36 +0000 | [diff] [blame] | 303 | def SDT_PCMPISTRI : SDTypeProfile<2, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, |
| 304 | SDTCisVT<2, v16i8>, SDTCisVT<3, v16i8>, |
| 305 | SDTCisVT<4, i8>]>; |
| 306 | def SDT_PCMPESTRI : SDTypeProfile<2, 5, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, |
| 307 | SDTCisVT<2, v16i8>, SDTCisVT<3, i32>, |
| 308 | SDTCisVT<4, v16i8>, SDTCisVT<5, i32>, |
| 309 | SDTCisVT<6, i8>]>; |
| 310 | |
| 311 | def X86pcmpistri : SDNode<"X86ISD::PCMPISTRI", SDT_PCMPISTRI>; |
| 312 | def X86pcmpestri : SDNode<"X86ISD::PCMPESTRI", SDT_PCMPESTRI>; |
| 313 | |
Elena Demikhovsky | 908dbf4 | 2014-12-11 15:02:24 +0000 | [diff] [blame] | 314 | def X86compress: SDNode<"X86ISD::COMPRESS", SDTypeProfile<1, 3, |
| 315 | [SDTCisSameAs<0, 2>, SDTCisSameAs<0, 3>, |
| 316 | SDTCisVec<3>, SDTCisVec<1>, SDTCisInt<1>]>, []>; |
Elena Demikhovsky | 72860c3 | 2014-12-15 10:03:52 +0000 | [diff] [blame] | 317 | def X86expand : SDNode<"X86ISD::EXPAND", SDTypeProfile<1, 3, |
| 318 | [SDTCisSameAs<0, 3>, |
| 319 | SDTCisVec<3>, SDTCisVec<1>, SDTCisInt<1>]>, []>; |
Elena Demikhovsky | 908dbf4 | 2014-12-11 15:02:24 +0000 | [diff] [blame] | 320 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 321 | //===----------------------------------------------------------------------===// |
| 322 | // SSE Complex Patterns |
| 323 | //===----------------------------------------------------------------------===// |
| 324 | |
| 325 | // These are 'extloads' from a scalar to the low element of a vector, zeroing |
| 326 | // the top elements. These are used for the SSE 'ss' and 'sd' instruction |
| 327 | // forms. |
| 328 | def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [], |
Chris Lattner | 0e023ea | 2010-09-21 20:31:19 +0000 | [diff] [blame] | 329 | [SDNPHasChain, SDNPMayLoad, SDNPMemOperand, |
| 330 | SDNPWantRoot]>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 331 | def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [], |
Chris Lattner | 0e023ea | 2010-09-21 20:31:19 +0000 | [diff] [blame] | 332 | [SDNPHasChain, SDNPMayLoad, SDNPMemOperand, |
| 333 | SDNPWantRoot]>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 334 | |
| 335 | def ssmem : Operand<v4f32> { |
| 336 | let PrintMethod = "printf32mem"; |
| 337 | let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm); |
Craig Topper | 6269f49 | 2013-08-26 00:39:04 +0000 | [diff] [blame] | 338 | let ParserMatchClass = X86Mem32AsmOperand; |
Benjamin Kramer | 9654eef | 2011-07-14 21:47:22 +0000 | [diff] [blame] | 339 | let OperandType = "OPERAND_MEMORY"; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 340 | } |
| 341 | def sdmem : Operand<v2f64> { |
| 342 | let PrintMethod = "printf64mem"; |
| 343 | let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm); |
Craig Topper | 6269f49 | 2013-08-26 00:39:04 +0000 | [diff] [blame] | 344 | let ParserMatchClass = X86Mem64AsmOperand; |
Benjamin Kramer | 9654eef | 2011-07-14 21:47:22 +0000 | [diff] [blame] | 345 | let OperandType = "OPERAND_MEMORY"; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 346 | } |
| 347 | |
| 348 | //===----------------------------------------------------------------------===// |
| 349 | // SSE pattern fragments |
| 350 | //===----------------------------------------------------------------------===// |
| 351 | |
Bruno Cardoso Lopes | 160be29 | 2010-08-13 20:39:01 +0000 | [diff] [blame] | 352 | // 128-bit load pattern fragments |
Craig Topper | 0d8e67a | 2012-01-24 03:03:17 +0000 | [diff] [blame] | 353 | // NOTE: all 128-bit integer vector loads are promoted to v2i64 |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 354 | def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>; |
| 355 | def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 356 | def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>; |
| 357 | |
Bruno Cardoso Lopes | 160be29 | 2010-08-13 20:39:01 +0000 | [diff] [blame] | 358 | // 256-bit load pattern fragments |
Craig Topper | 0d8e67a | 2012-01-24 03:03:17 +0000 | [diff] [blame] | 359 | // NOTE: all 256-bit integer vector loads are promoted to v4i64 |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 360 | def loadv8f32 : PatFrag<(ops node:$ptr), (v8f32 (load node:$ptr))>; |
| 361 | def loadv4f64 : PatFrag<(ops node:$ptr), (v4f64 (load node:$ptr))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 362 | def loadv4i64 : PatFrag<(ops node:$ptr), (v4i64 (load node:$ptr))>; |
| 363 | |
Elena Demikhovsky | cf5b145 | 2013-08-11 07:55:09 +0000 | [diff] [blame] | 364 | // 512-bit load pattern fragments |
| 365 | def loadv16f32 : PatFrag<(ops node:$ptr), (v16f32 (load node:$ptr))>; |
| 366 | def loadv8f64 : PatFrag<(ops node:$ptr), (v8f64 (load node:$ptr))>; |
Robert Khasanov | 7ca7df0 | 2014-08-04 14:35:15 +0000 | [diff] [blame] | 367 | def loadv64i8 : PatFrag<(ops node:$ptr), (v64i8 (load node:$ptr))>; |
| 368 | def loadv32i16 : PatFrag<(ops node:$ptr), (v32i16 (load node:$ptr))>; |
Elena Demikhovsky | 1f3ed41 | 2013-10-22 09:19:28 +0000 | [diff] [blame] | 369 | def loadv16i32 : PatFrag<(ops node:$ptr), (v16i32 (load node:$ptr))>; |
Elena Demikhovsky | cf5b145 | 2013-08-11 07:55:09 +0000 | [diff] [blame] | 370 | def loadv8i64 : PatFrag<(ops node:$ptr), (v8i64 (load node:$ptr))>; |
| 371 | |
| 372 | // 128-/256-/512-bit extload pattern fragments |
Michael Liao | 400f7ef | 2012-09-10 18:33:51 +0000 | [diff] [blame] | 373 | def extloadv2f32 : PatFrag<(ops node:$ptr), (v2f64 (extloadvf32 node:$ptr))>; |
| 374 | def extloadv4f32 : PatFrag<(ops node:$ptr), (v4f64 (extloadvf32 node:$ptr))>; |
Elena Demikhovsky | cf5b145 | 2013-08-11 07:55:09 +0000 | [diff] [blame] | 375 | def extloadv8f32 : PatFrag<(ops node:$ptr), (v8f64 (extloadvf32 node:$ptr))>; |
Michael Liao | 400f7ef | 2012-09-10 18:33:51 +0000 | [diff] [blame] | 376 | |
Sanjay Patel | b811c1d | 2015-02-17 20:08:21 +0000 | [diff] [blame] | 377 | // These are needed to match a scalar load that is used in a vector-only |
| 378 | // math instruction such as the FP logical ops: andps, andnps, orps, xorps. |
| 379 | // The memory operand is required to be a 128-bit load, so it must be converted |
| 380 | // from a vector to a scalar. |
| 381 | def loadf32_128 : PatFrag<(ops node:$ptr), |
| 382 | (f32 (vector_extract (loadv4f32 node:$ptr), (iPTR 0)))>; |
| 383 | def loadf64_128 : PatFrag<(ops node:$ptr), |
| 384 | (f64 (vector_extract (loadv2f64 node:$ptr), (iPTR 0)))>; |
| 385 | |
Bruno Cardoso Lopes | 03d6002 | 2011-09-13 19:33:03 +0000 | [diff] [blame] | 386 | // Like 'store', but always requires 128-bit vector alignment. |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 387 | def alignedstore : PatFrag<(ops node:$val, node:$ptr), |
| 388 | (store node:$val, node:$ptr), [{ |
| 389 | return cast<StoreSDNode>(N)->getAlignment() >= 16; |
| 390 | }]>; |
| 391 | |
Bruno Cardoso Lopes | 03d6002 | 2011-09-13 19:33:03 +0000 | [diff] [blame] | 392 | // Like 'store', but always requires 256-bit vector alignment. |
| 393 | def alignedstore256 : PatFrag<(ops node:$val, node:$ptr), |
| 394 | (store node:$val, node:$ptr), [{ |
| 395 | return cast<StoreSDNode>(N)->getAlignment() >= 32; |
| 396 | }]>; |
| 397 | |
Elena Demikhovsky | cf5b145 | 2013-08-11 07:55:09 +0000 | [diff] [blame] | 398 | // Like 'store', but always requires 512-bit vector alignment. |
| 399 | def alignedstore512 : PatFrag<(ops node:$val, node:$ptr), |
| 400 | (store node:$val, node:$ptr), [{ |
| 401 | return cast<StoreSDNode>(N)->getAlignment() >= 64; |
| 402 | }]>; |
| 403 | |
Bruno Cardoso Lopes | 03d6002 | 2011-09-13 19:33:03 +0000 | [diff] [blame] | 404 | // Like 'load', but always requires 128-bit vector alignment. |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 405 | def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{ |
| 406 | return cast<LoadSDNode>(N)->getAlignment() >= 16; |
| 407 | }]>; |
| 408 | |
Chad Rosier | a281afc | 2012-03-09 02:00:48 +0000 | [diff] [blame] | 409 | // Like 'X86vzload', but always requires 128-bit vector alignment. |
| 410 | def alignedX86vzload : PatFrag<(ops node:$ptr), (X86vzload node:$ptr), [{ |
| 411 | return cast<MemSDNode>(N)->getAlignment() >= 16; |
| 412 | }]>; |
| 413 | |
Bruno Cardoso Lopes | 03d6002 | 2011-09-13 19:33:03 +0000 | [diff] [blame] | 414 | // Like 'load', but always requires 256-bit vector alignment. |
| 415 | def alignedload256 : PatFrag<(ops node:$ptr), (load node:$ptr), [{ |
| 416 | return cast<LoadSDNode>(N)->getAlignment() >= 32; |
| 417 | }]>; |
| 418 | |
Elena Demikhovsky | cf5b145 | 2013-08-11 07:55:09 +0000 | [diff] [blame] | 419 | // Like 'load', but always requires 512-bit vector alignment. |
| 420 | def alignedload512 : PatFrag<(ops node:$ptr), (load node:$ptr), [{ |
| 421 | return cast<LoadSDNode>(N)->getAlignment() >= 64; |
| 422 | }]>; |
| 423 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 424 | def alignedloadfsf32 : PatFrag<(ops node:$ptr), |
| 425 | (f32 (alignedload node:$ptr))>; |
| 426 | def alignedloadfsf64 : PatFrag<(ops node:$ptr), |
| 427 | (f64 (alignedload node:$ptr))>; |
Bruno Cardoso Lopes | 160be29 | 2010-08-13 20:39:01 +0000 | [diff] [blame] | 428 | |
| 429 | // 128-bit aligned load pattern fragments |
Craig Topper | 0d8e67a | 2012-01-24 03:03:17 +0000 | [diff] [blame] | 430 | // NOTE: all 128-bit integer vector loads are promoted to v2i64 |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 431 | def alignedloadv4f32 : PatFrag<(ops node:$ptr), |
| 432 | (v4f32 (alignedload node:$ptr))>; |
| 433 | def alignedloadv2f64 : PatFrag<(ops node:$ptr), |
| 434 | (v2f64 (alignedload node:$ptr))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 435 | def alignedloadv2i64 : PatFrag<(ops node:$ptr), |
| 436 | (v2i64 (alignedload node:$ptr))>; |
| 437 | |
Bruno Cardoso Lopes | 160be29 | 2010-08-13 20:39:01 +0000 | [diff] [blame] | 438 | // 256-bit aligned load pattern fragments |
Craig Topper | 0d8e67a | 2012-01-24 03:03:17 +0000 | [diff] [blame] | 439 | // NOTE: all 256-bit integer vector loads are promoted to v4i64 |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 440 | def alignedloadv8f32 : PatFrag<(ops node:$ptr), |
Bruno Cardoso Lopes | 03d6002 | 2011-09-13 19:33:03 +0000 | [diff] [blame] | 441 | (v8f32 (alignedload256 node:$ptr))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 442 | def alignedloadv4f64 : PatFrag<(ops node:$ptr), |
Bruno Cardoso Lopes | 03d6002 | 2011-09-13 19:33:03 +0000 | [diff] [blame] | 443 | (v4f64 (alignedload256 node:$ptr))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 444 | def alignedloadv4i64 : PatFrag<(ops node:$ptr), |
Bruno Cardoso Lopes | 03d6002 | 2011-09-13 19:33:03 +0000 | [diff] [blame] | 445 | (v4i64 (alignedload256 node:$ptr))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 446 | |
Elena Demikhovsky | cf5b145 | 2013-08-11 07:55:09 +0000 | [diff] [blame] | 447 | // 512-bit aligned load pattern fragments |
| 448 | def alignedloadv16f32 : PatFrag<(ops node:$ptr), |
| 449 | (v16f32 (alignedload512 node:$ptr))>; |
Elena Demikhovsky | 3ce8dbb | 2013-08-18 13:08:57 +0000 | [diff] [blame] | 450 | def alignedloadv16i32 : PatFrag<(ops node:$ptr), |
| 451 | (v16i32 (alignedload512 node:$ptr))>; |
Elena Demikhovsky | cf5b145 | 2013-08-11 07:55:09 +0000 | [diff] [blame] | 452 | def alignedloadv8f64 : PatFrag<(ops node:$ptr), |
| 453 | (v8f64 (alignedload512 node:$ptr))>; |
Elena Demikhovsky | cf5b145 | 2013-08-11 07:55:09 +0000 | [diff] [blame] | 454 | def alignedloadv8i64 : PatFrag<(ops node:$ptr), |
| 455 | (v8i64 (alignedload512 node:$ptr))>; |
| 456 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 457 | // Like 'load', but uses special alignment checks suitable for use in |
| 458 | // memory operands in most SSE instructions, which are required to |
| 459 | // be naturally aligned on some targets but not on others. If the subtarget |
| 460 | // allows unaligned accesses, match any load, though this may require |
| 461 | // setting a feature bit in the processor (on startup, for example). |
| 462 | // Opteron 10h and later implement such a feature. |
| 463 | def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{ |
Sanjay Patel | ffd039b | 2015-02-03 17:13:04 +0000 | [diff] [blame] | 464 | return Subtarget->hasSSEUnalignedMem() |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 465 | || cast<LoadSDNode>(N)->getAlignment() >= 16; |
| 466 | }]>; |
| 467 | |
| 468 | def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>; |
| 469 | def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>; |
Bruno Cardoso Lopes | 160be29 | 2010-08-13 20:39:01 +0000 | [diff] [blame] | 470 | |
| 471 | // 128-bit memop pattern fragments |
Craig Topper | 0d8e67a | 2012-01-24 03:03:17 +0000 | [diff] [blame] | 472 | // NOTE: all 128-bit integer vector loads are promoted to v2i64 |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 473 | def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>; |
| 474 | def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 475 | def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 476 | |
Sanjay Patel | b811c1d | 2015-02-17 20:08:21 +0000 | [diff] [blame] | 477 | // These are needed to match a scalar memop that is used in a vector-only |
| 478 | // math instruction such as the FP logical ops: andps, andnps, orps, xorps. |
| 479 | // The memory operand is required to be a 128-bit load, so it must be converted |
| 480 | // from a vector to a scalar. |
| 481 | def memopfsf32_128 : PatFrag<(ops node:$ptr), |
| 482 | (f32 (vector_extract (memopv4f32 node:$ptr), (iPTR 0)))>; |
| 483 | def memopfsf64_128 : PatFrag<(ops node:$ptr), |
| 484 | (f64 (vector_extract (memopv2f64 node:$ptr), (iPTR 0)))>; |
| 485 | |
| 486 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 487 | // SSSE3 uses MMX registers for some instructions. They aren't aligned on a |
| 488 | // 16-byte boundary. |
| 489 | // FIXME: 8 byte alignment for mmx reads is not required |
| 490 | def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{ |
| 491 | return cast<LoadSDNode>(N)->getAlignment() >= 8; |
| 492 | }]>; |
| 493 | |
Dale Johannesen | dd224d2 | 2010-09-30 23:57:10 +0000 | [diff] [blame] | 494 | def memopmmx : PatFrag<(ops node:$ptr), (x86mmx (memop64 node:$ptr))>; |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 495 | |
| 496 | // MOVNT Support |
| 497 | // Like 'store', but requires the non-temporal bit to be set |
| 498 | def nontemporalstore : PatFrag<(ops node:$val, node:$ptr), |
| 499 | (st node:$val, node:$ptr), [{ |
| 500 | if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) |
| 501 | return ST->isNonTemporal(); |
| 502 | return false; |
| 503 | }]>; |
| 504 | |
| 505 | def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr), |
Bill Wendling | ea6397f | 2012-07-19 00:11:40 +0000 | [diff] [blame] | 506 | (st node:$val, node:$ptr), [{ |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 507 | if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) |
| 508 | return ST->isNonTemporal() && !ST->isTruncatingStore() && |
| 509 | ST->getAddressingMode() == ISD::UNINDEXED && |
| 510 | ST->getAlignment() >= 16; |
| 511 | return false; |
| 512 | }]>; |
| 513 | |
| 514 | def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr), |
Bill Wendling | ea6397f | 2012-07-19 00:11:40 +0000 | [diff] [blame] | 515 | (st node:$val, node:$ptr), [{ |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 516 | if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) |
| 517 | return ST->isNonTemporal() && |
| 518 | ST->getAlignment() < 16; |
| 519 | return false; |
| 520 | }]>; |
| 521 | |
Bruno Cardoso Lopes | 160be29 | 2010-08-13 20:39:01 +0000 | [diff] [blame] | 522 | // 128-bit bitconvert pattern fragments |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 523 | def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>; |
| 524 | def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>; |
| 525 | def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>; |
| 526 | def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>; |
| 527 | def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>; |
| 528 | def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>; |
| 529 | |
Bruno Cardoso Lopes | 160be29 | 2010-08-13 20:39:01 +0000 | [diff] [blame] | 530 | // 256-bit bitconvert pattern fragments |
Craig Topper | 682b850 | 2011-11-02 04:42:13 +0000 | [diff] [blame] | 531 | def bc_v32i8 : PatFrag<(ops node:$in), (v32i8 (bitconvert node:$in))>; |
| 532 | def bc_v16i16 : PatFrag<(ops node:$in), (v16i16 (bitconvert node:$in))>; |
Bruno Cardoso Lopes | e3acfd4 | 2010-07-21 23:53:50 +0000 | [diff] [blame] | 533 | def bc_v8i32 : PatFrag<(ops node:$in), (v8i32 (bitconvert node:$in))>; |
Bruno Cardoso Lopes | 1021b4a | 2011-07-13 01:15:33 +0000 | [diff] [blame] | 534 | def bc_v4i64 : PatFrag<(ops node:$in), (v4i64 (bitconvert node:$in))>; |
Elena Demikhovsky | 3629b4a | 2014-01-06 08:45:54 +0000 | [diff] [blame] | 535 | def bc_v8f32 : PatFrag<(ops node:$in), (v8f32 (bitconvert node:$in))>; |
Bruno Cardoso Lopes | e3acfd4 | 2010-07-21 23:53:50 +0000 | [diff] [blame] | 536 | |
Craig Topper | 8c92962 | 2013-08-16 06:07:34 +0000 | [diff] [blame] | 537 | // 512-bit bitconvert pattern fragments |
| 538 | def bc_v16i32 : PatFrag<(ops node:$in), (v16i32 (bitconvert node:$in))>; |
| 539 | def bc_v8i64 : PatFrag<(ops node:$in), (v8i64 (bitconvert node:$in))>; |
Elena Demikhovsky | de3f751 | 2014-01-01 15:12:34 +0000 | [diff] [blame] | 540 | def bc_v8f64 : PatFrag<(ops node:$in), (v8f64 (bitconvert node:$in))>; |
| 541 | def bc_v16f32 : PatFrag<(ops node:$in), (v16f32 (bitconvert node:$in))>; |
Craig Topper | 8c92962 | 2013-08-16 06:07:34 +0000 | [diff] [blame] | 542 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 543 | def vzmovl_v2i64 : PatFrag<(ops node:$src), |
| 544 | (bitconvert (v2i64 (X86vzmovl |
| 545 | (v2i64 (scalar_to_vector (loadi64 node:$src))))))>; |
| 546 | def vzmovl_v4i32 : PatFrag<(ops node:$src), |
| 547 | (bitconvert (v4i32 (X86vzmovl |
| 548 | (v4i32 (scalar_to_vector (loadi32 node:$src))))))>; |
| 549 | |
| 550 | def vzload_v2i64 : PatFrag<(ops node:$src), |
| 551 | (bitconvert (v2i64 (X86vzload node:$src)))>; |
| 552 | |
| 553 | |
| 554 | def fp32imm0 : PatLeaf<(f32 fpimm), [{ |
| 555 | return N->isExactlyValue(+0.0); |
| 556 | }]>; |
| 557 | |
Elena Demikhovsky | de3f751 | 2014-01-01 15:12:34 +0000 | [diff] [blame] | 558 | def I8Imm : SDNodeXForm<imm, [{ |
| 559 | // Transformation function: get the low 8 bits. |
| 560 | return getI8Imm((uint8_t)N->getZExtValue()); |
| 561 | }]>; |
| 562 | |
| 563 | def FROUND_NO_EXC : ImmLeaf<i32, [{ return Imm == 8; }]>; |
Adam Nemet | 50b83f0 | 2014-08-14 17:13:26 +0000 | [diff] [blame] | 564 | def FROUND_CURRENT : ImmLeaf<i32, [{ |
| 565 | return Imm == X86::STATIC_ROUNDING::CUR_DIRECTION; |
| 566 | }]>; |
Elena Demikhovsky | de3f751 | 2014-01-01 15:12:34 +0000 | [diff] [blame] | 567 | |
David Greene | 03264ef | 2010-07-12 23:41:28 +0000 | [diff] [blame] | 568 | // BYTE_imm - Transform bit immediates into byte immediates. |
| 569 | def BYTE_imm : SDNodeXForm<imm, [{ |
| 570 | // Transformation function: imm >> 3 |
| 571 | return getI32Imm(N->getZExtValue() >> 3); |
| 572 | }]>; |
| 573 | |
Elena Demikhovsky | 67b05fc | 2013-07-31 11:35:14 +0000 | [diff] [blame] | 574 | // EXTRACT_get_vextract128_imm xform function: convert extract_subvector index |
| 575 | // to VEXTRACTF128/VEXTRACTI128 imm. |
| 576 | def EXTRACT_get_vextract128_imm : SDNodeXForm<extract_subvector, [{ |
| 577 | return getI8Imm(X86::getExtractVEXTRACT128Immediate(N)); |
David Greene | c4da110 | 2011-02-03 15:50:00 +0000 | [diff] [blame] | 578 | }]>; |
| 579 | |
Elena Demikhovsky | 67b05fc | 2013-07-31 11:35:14 +0000 | [diff] [blame] | 580 | // INSERT_get_vinsert128_imm xform function: convert insert_subvector index to |
| 581 | // VINSERTF128/VINSERTI128 imm. |
| 582 | def INSERT_get_vinsert128_imm : SDNodeXForm<insert_subvector, [{ |
| 583 | return getI8Imm(X86::getInsertVINSERT128Immediate(N)); |
David Greene | 653f1ee | 2011-02-04 16:08:29 +0000 | [diff] [blame] | 584 | }]>; |
| 585 | |
Elena Demikhovsky | 67b05fc | 2013-07-31 11:35:14 +0000 | [diff] [blame] | 586 | // EXTRACT_get_vextract256_imm xform function: convert extract_subvector index |
| 587 | // to VEXTRACTF64x4 imm. |
| 588 | def EXTRACT_get_vextract256_imm : SDNodeXForm<extract_subvector, [{ |
| 589 | return getI8Imm(X86::getExtractVEXTRACT256Immediate(N)); |
| 590 | }]>; |
| 591 | |
| 592 | // INSERT_get_vinsert256_imm xform function: convert insert_subvector index to |
| 593 | // VINSERTF64x4 imm. |
| 594 | def INSERT_get_vinsert256_imm : SDNodeXForm<insert_subvector, [{ |
| 595 | return getI8Imm(X86::getInsertVINSERT256Immediate(N)); |
| 596 | }]>; |
| 597 | |
| 598 | def vextract128_extract : PatFrag<(ops node:$bigvec, node:$index), |
David Greene | c4da110 | 2011-02-03 15:50:00 +0000 | [diff] [blame] | 599 | (extract_subvector node:$bigvec, |
| 600 | node:$index), [{ |
Elena Demikhovsky | 67b05fc | 2013-07-31 11:35:14 +0000 | [diff] [blame] | 601 | return X86::isVEXTRACT128Index(N); |
| 602 | }], EXTRACT_get_vextract128_imm>; |
David Greene | 653f1ee | 2011-02-04 16:08:29 +0000 | [diff] [blame] | 603 | |
Elena Demikhovsky | 67b05fc | 2013-07-31 11:35:14 +0000 | [diff] [blame] | 604 | def vinsert128_insert : PatFrag<(ops node:$bigvec, node:$smallvec, |
David Greene | 653f1ee | 2011-02-04 16:08:29 +0000 | [diff] [blame] | 605 | node:$index), |
| 606 | (insert_subvector node:$bigvec, node:$smallvec, |
| 607 | node:$index), [{ |
Elena Demikhovsky | 67b05fc | 2013-07-31 11:35:14 +0000 | [diff] [blame] | 608 | return X86::isVINSERT128Index(N); |
| 609 | }], INSERT_get_vinsert128_imm>; |
| 610 | |
| 611 | |
| 612 | def vextract256_extract : PatFrag<(ops node:$bigvec, node:$index), |
| 613 | (extract_subvector node:$bigvec, |
| 614 | node:$index), [{ |
| 615 | return X86::isVEXTRACT256Index(N); |
| 616 | }], EXTRACT_get_vextract256_imm>; |
| 617 | |
| 618 | def vinsert256_insert : PatFrag<(ops node:$bigvec, node:$smallvec, |
| 619 | node:$index), |
| 620 | (insert_subvector node:$bigvec, node:$smallvec, |
| 621 | node:$index), [{ |
| 622 | return X86::isVINSERT256Index(N); |
| 623 | }], INSERT_get_vinsert256_imm>; |
Bruno Cardoso Lopes | 123dff0 | 2011-07-25 23:05:25 +0000 | [diff] [blame] | 624 | |