[SystemZ] Add NC, OC and XC
[oota-llvm.git] / lib / Target / SystemZ / SystemZOperators.td
1 //===-- SystemZOperators.td - SystemZ-specific operators ------*- tblgen-*-===//
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.
7 //
8 //===----------------------------------------------------------------------===//
9
10 //===----------------------------------------------------------------------===//
11 // Type profiles
12 //===----------------------------------------------------------------------===//
13 def SDT_CallSeqStart        : SDCallSeqStart<[SDTCisVT<0, i64>]>;
14 def SDT_CallSeqEnd          : SDCallSeqEnd<[SDTCisVT<0, i64>,
15                                             SDTCisVT<1, i64>]>;
16 def SDT_ZCall               : SDTypeProfile<0, -1, [SDTCisPtrTy<0>]>;
17 def SDT_ZCmp                : SDTypeProfile<0, 2, [SDTCisSameAs<0, 1>]>;
18 def SDT_ZBRCCMask           : SDTypeProfile<0, 3,
19                                             [SDTCisVT<0, i8>,
20                                              SDTCisVT<1, i8>,
21                                              SDTCisVT<2, OtherVT>]>;
22 def SDT_ZSelectCCMask       : SDTypeProfile<1, 4,
23                                             [SDTCisSameAs<0, 1>,
24                                              SDTCisSameAs<1, 2>,
25                                              SDTCisVT<3, i8>,
26                                              SDTCisVT<4, i8>]>;
27 def SDT_ZWrapPtr            : SDTypeProfile<1, 1,
28                                             [SDTCisSameAs<0, 1>,
29                                              SDTCisPtrTy<0>]>;
30 def SDT_ZAdjDynAlloc        : SDTypeProfile<1, 0, [SDTCisVT<0, i64>]>;
31 def SDT_ZExtractAccess      : SDTypeProfile<1, 1,
32                                             [SDTCisVT<0, i32>,
33                                              SDTCisVT<1, i8>]>;
34 def SDT_ZGR128Binary32      : SDTypeProfile<1, 2,
35                                             [SDTCisVT<0, untyped>,
36                                              SDTCisVT<1, untyped>,
37                                              SDTCisVT<2, i32>]>;
38 def SDT_ZGR128Binary64      : SDTypeProfile<1, 2,
39                                             [SDTCisVT<0, untyped>,
40                                              SDTCisVT<1, untyped>,
41                                              SDTCisVT<2, i64>]>;
42 def SDT_ZAtomicLoadBinaryW  : SDTypeProfile<1, 5,
43                                             [SDTCisVT<0, i32>,
44                                              SDTCisPtrTy<1>,
45                                              SDTCisVT<2, i32>,
46                                              SDTCisVT<3, i32>,
47                                              SDTCisVT<4, i32>,
48                                              SDTCisVT<5, i32>]>;
49 def SDT_ZAtomicCmpSwapW     : SDTypeProfile<1, 6,
50                                             [SDTCisVT<0, i32>,
51                                              SDTCisPtrTy<1>,
52                                              SDTCisVT<2, i32>,
53                                              SDTCisVT<3, i32>,
54                                              SDTCisVT<4, i32>,
55                                              SDTCisVT<5, i32>,
56                                              SDTCisVT<6, i32>]>;
57 def SDT_ZMemMemLength       : SDTypeProfile<0, 3,
58                                             [SDTCisPtrTy<0>,
59                                              SDTCisPtrTy<1>,
60                                              SDTCisVT<2, i64>]>;
61 def SDT_ZMemMemLoop         : SDTypeProfile<0, 4,
62                                             [SDTCisPtrTy<0>,
63                                              SDTCisPtrTy<1>,
64                                              SDTCisVT<2, i64>,
65                                              SDTCisVT<3, i64>]>;
66 def SDT_ZString             : SDTypeProfile<1, 3,
67                                             [SDTCisPtrTy<0>,
68                                              SDTCisPtrTy<1>,
69                                              SDTCisPtrTy<2>,
70                                              SDTCisVT<3, i32>]>;
71 def SDT_ZI32Intrinsic       : SDTypeProfile<1, 0, [SDTCisVT<0, i32>]>;
72 def SDT_ZPrefetch           : SDTypeProfile<0, 2,
73                                             [SDTCisVT<0, i8>,
74                                              SDTCisPtrTy<1>]>;
75
76 //===----------------------------------------------------------------------===//
77 // Node definitions
78 //===----------------------------------------------------------------------===//
79
80 // These are target-independent nodes, but have target-specific formats.
81 def callseq_start       : SDNode<"ISD::CALLSEQ_START", SDT_CallSeqStart,
82                                  [SDNPHasChain, SDNPSideEffect, SDNPOutGlue]>;
83 def callseq_end         : SDNode<"ISD::CALLSEQ_END",   SDT_CallSeqEnd,
84                                  [SDNPHasChain, SDNPSideEffect, SDNPOptInGlue,
85                                   SDNPOutGlue]>;
86
87 // Nodes for SystemZISD::*.  See SystemZISelLowering.h for more details.
88 def z_retflag           : SDNode<"SystemZISD::RET_FLAG", SDTNone,
89                                  [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
90 def z_call              : SDNode<"SystemZISD::CALL", SDT_ZCall,
91                                  [SDNPHasChain, SDNPOutGlue, SDNPOptInGlue,
92                                   SDNPVariadic]>;
93 def z_sibcall           : SDNode<"SystemZISD::SIBCALL", SDT_ZCall,
94                                  [SDNPHasChain, SDNPOutGlue, SDNPOptInGlue,
95                                   SDNPVariadic]>;
96 def z_pcrel_wrapper     : SDNode<"SystemZISD::PCREL_WRAPPER", SDT_ZWrapPtr, []>;
97 def z_cmp               : SDNode<"SystemZISD::CMP", SDT_ZCmp, [SDNPOutGlue]>;
98 def z_ucmp              : SDNode<"SystemZISD::UCMP", SDT_ZCmp, [SDNPOutGlue]>;
99 def z_tm                : SDNode<"SystemZISD::TM", SDT_ZCmp, [SDNPOutGlue]>;
100 def z_br_ccmask         : SDNode<"SystemZISD::BR_CCMASK", SDT_ZBRCCMask,
101                                  [SDNPHasChain, SDNPInGlue]>;
102 def z_select_ccmask     : SDNode<"SystemZISD::SELECT_CCMASK", SDT_ZSelectCCMask,
103                                  [SDNPInGlue]>;
104 def z_adjdynalloc       : SDNode<"SystemZISD::ADJDYNALLOC", SDT_ZAdjDynAlloc>;
105 def z_extract_access    : SDNode<"SystemZISD::EXTRACT_ACCESS",
106                                  SDT_ZExtractAccess>;
107 def z_umul_lohi64       : SDNode<"SystemZISD::UMUL_LOHI64", SDT_ZGR128Binary64>;
108 def z_sdivrem32         : SDNode<"SystemZISD::SDIVREM32", SDT_ZGR128Binary32>;
109 def z_sdivrem64         : SDNode<"SystemZISD::SDIVREM64", SDT_ZGR128Binary64>;
110 def z_udivrem32         : SDNode<"SystemZISD::UDIVREM32", SDT_ZGR128Binary32>;
111 def z_udivrem64         : SDNode<"SystemZISD::UDIVREM64", SDT_ZGR128Binary64>;
112
113 class AtomicWOp<string name, SDTypeProfile profile = SDT_ZAtomicLoadBinaryW>
114   : SDNode<"SystemZISD::"##name, profile,
115            [SDNPHasChain, SDNPMayStore, SDNPMayLoad, SDNPMemOperand]>;
116
117 def z_atomic_swapw      : AtomicWOp<"ATOMIC_SWAPW">;
118 def z_atomic_loadw_add  : AtomicWOp<"ATOMIC_LOADW_ADD">;
119 def z_atomic_loadw_sub  : AtomicWOp<"ATOMIC_LOADW_SUB">;
120 def z_atomic_loadw_and  : AtomicWOp<"ATOMIC_LOADW_AND">;
121 def z_atomic_loadw_or   : AtomicWOp<"ATOMIC_LOADW_OR">;
122 def z_atomic_loadw_xor  : AtomicWOp<"ATOMIC_LOADW_XOR">;
123 def z_atomic_loadw_nand : AtomicWOp<"ATOMIC_LOADW_NAND">;
124 def z_atomic_loadw_min  : AtomicWOp<"ATOMIC_LOADW_MIN">;
125 def z_atomic_loadw_max  : AtomicWOp<"ATOMIC_LOADW_MAX">;
126 def z_atomic_loadw_umin : AtomicWOp<"ATOMIC_LOADW_UMIN">;
127 def z_atomic_loadw_umax : AtomicWOp<"ATOMIC_LOADW_UMAX">;
128 def z_atomic_cmp_swapw  : AtomicWOp<"ATOMIC_CMP_SWAPW", SDT_ZAtomicCmpSwapW>;
129
130 def z_mvc               : SDNode<"SystemZISD::MVC", SDT_ZMemMemLength,
131                                  [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
132 def z_mvc_loop          : SDNode<"SystemZISD::MVC_LOOP", SDT_ZMemMemLoop,
133                                  [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
134 def z_nc                : SDNode<"SystemZISD::NC", SDT_ZMemMemLength,
135                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
136 def z_nc_loop           : SDNode<"SystemZISD::NC_LOOP", SDT_ZMemMemLoop,
137                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
138 def z_oc                : SDNode<"SystemZISD::OC", SDT_ZMemMemLength,
139                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
140 def z_oc_loop           : SDNode<"SystemZISD::OC_LOOP", SDT_ZMemMemLoop,
141                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
142 def z_xc                : SDNode<"SystemZISD::XC", SDT_ZMemMemLength,
143                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
144 def z_xc_loop           : SDNode<"SystemZISD::XC_LOOP", SDT_ZMemMemLoop,
145                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
146 def z_clc               : SDNode<"SystemZISD::CLC", SDT_ZMemMemLength,
147                                  [SDNPHasChain, SDNPOutGlue, SDNPMayLoad]>;
148 def z_clc_loop          : SDNode<"SystemZISD::CLC_LOOP", SDT_ZMemMemLoop,
149                                  [SDNPHasChain, SDNPOutGlue, SDNPMayLoad]>;
150 def z_strcmp            : SDNode<"SystemZISD::STRCMP", SDT_ZString,
151                                  [SDNPHasChain, SDNPOutGlue, SDNPMayLoad]>;
152 def z_stpcpy            : SDNode<"SystemZISD::STPCPY", SDT_ZString,
153                                  [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
154 def z_search_string     : SDNode<"SystemZISD::SEARCH_STRING", SDT_ZString,
155                                  [SDNPHasChain, SDNPOutGlue, SDNPMayLoad]>;
156 def z_ipm               : SDNode<"SystemZISD::IPM", SDT_ZI32Intrinsic,
157                                  [SDNPInGlue]>;
158 def z_prefetch          : SDNode<"SystemZISD::PREFETCH", SDT_ZPrefetch,
159                                  [SDNPHasChain, SDNPMayLoad, SDNPMayStore,
160                                   SDNPMemOperand]>;
161
162 //===----------------------------------------------------------------------===//
163 // Pattern fragments
164 //===----------------------------------------------------------------------===//
165
166 // Register sign-extend operations.  Sub-32-bit values are represented as i32s.
167 def sext8  : PatFrag<(ops node:$src), (sext_inreg node:$src, i8)>;
168 def sext16 : PatFrag<(ops node:$src), (sext_inreg node:$src, i16)>;
169 def sext32 : PatFrag<(ops node:$src), (sext (i32 node:$src))>;
170
171 // Register zero-extend operations.  Sub-32-bit values are represented as i32s.
172 def zext8  : PatFrag<(ops node:$src), (and node:$src, 0xff)>;
173 def zext16 : PatFrag<(ops node:$src), (and node:$src, 0xffff)>;
174 def zext32 : PatFrag<(ops node:$src), (zext (i32 node:$src))>;
175
176 // Typed floating-point loads.
177 def loadf32 : PatFrag<(ops node:$src), (f32 (load node:$src))>;
178 def loadf64 : PatFrag<(ops node:$src), (f64 (load node:$src))>;
179
180 // Extending loads in which the extension type doesn't matter.
181 def anyextload : PatFrag<(ops node:$ptr), (unindexedload node:$ptr), [{
182   return cast<LoadSDNode>(N)->getExtensionType() != ISD::NON_EXTLOAD;
183 }]>;
184 def anyextloadi8 : PatFrag<(ops node:$ptr), (anyextload node:$ptr), [{
185   return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i8;
186 }]>;
187 def anyextloadi16 : PatFrag<(ops node:$ptr), (anyextload node:$ptr), [{
188   return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i16;
189 }]>;
190 def anyextloadi32 : PatFrag<(ops node:$ptr), (anyextload node:$ptr), [{
191   return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i32;
192 }]>;
193
194 // Aligned loads.
195 class AlignedLoad<SDPatternOperator load>
196   : PatFrag<(ops node:$addr), (load node:$addr), [{
197   LoadSDNode *Load = cast<LoadSDNode>(N);
198   return Load->getAlignment() >= Load->getMemoryVT().getStoreSize();
199 }]>;
200 def aligned_load        : AlignedLoad<load>;
201 def aligned_sextloadi16 : AlignedLoad<sextloadi16>;
202 def aligned_sextloadi32 : AlignedLoad<sextloadi32>;
203 def aligned_zextloadi16 : AlignedLoad<zextloadi16>;
204 def aligned_zextloadi32 : AlignedLoad<zextloadi32>;
205
206 // Aligned stores.
207 class AlignedStore<SDPatternOperator store>
208   : PatFrag<(ops node:$src, node:$addr), (store node:$src, node:$addr), [{
209   StoreSDNode *Store = cast<StoreSDNode>(N);
210   return Store->getAlignment() >= Store->getMemoryVT().getStoreSize();
211 }]>;
212 def aligned_store         : AlignedStore<store>;
213 def aligned_truncstorei16 : AlignedStore<truncstorei16>;
214 def aligned_truncstorei32 : AlignedStore<truncstorei32>;
215
216 // Non-volatile loads.  Used for instructions that might access the storage
217 // location multiple times.
218 class NonvolatileLoad<SDPatternOperator load>
219   : PatFrag<(ops node:$addr), (load node:$addr), [{
220   LoadSDNode *Load = cast<LoadSDNode>(N);
221   return !Load->isVolatile();
222 }]>;
223 def nonvolatile_load          : NonvolatileLoad<load>;
224 def nonvolatile_anyextloadi8  : NonvolatileLoad<anyextloadi8>;
225 def nonvolatile_anyextloadi16 : NonvolatileLoad<anyextloadi16>;
226 def nonvolatile_anyextloadi32 : NonvolatileLoad<anyextloadi32>;
227
228 // Non-volatile stores.
229 class NonvolatileStore<SDPatternOperator store>
230   : PatFrag<(ops node:$src, node:$addr), (store node:$src, node:$addr), [{
231   StoreSDNode *Store = cast<StoreSDNode>(N);
232   return !Store->isVolatile();
233 }]>;
234 def nonvolatile_store         : NonvolatileStore<store>;
235 def nonvolatile_truncstorei8  : NonvolatileStore<truncstorei8>;
236 def nonvolatile_truncstorei16 : NonvolatileStore<truncstorei16>;
237 def nonvolatile_truncstorei32 : NonvolatileStore<truncstorei32>;
238
239 // A store of a load that can be implemented using MVC.
240 def mvc_store : PatFrag<(ops node:$value, node:$addr),
241                         (unindexedstore node:$value, node:$addr),
242                         [{ return storeLoadCanUseMVC(N); }]>;
243
244 // Binary read-modify-write operations on memory in which the other
245 // operand is also memory and for which block operations like NC can
246 // be used.  There are two patterns for each operator, depending on
247 // which operand contains the "other" load.
248 multiclass block_op<SDPatternOperator operator> {
249   def "1" : PatFrag<(ops node:$value, node:$addr),
250                     (unindexedstore (operator node:$value,
251                                               (unindexedload node:$addr)),
252                                     node:$addr),
253                     [{ return storeLoadCanUseBlockBinary(N, 0); }]>;
254   def "2" : PatFrag<(ops node:$value, node:$addr),
255                     (unindexedstore (operator (unindexedload node:$addr),
256                                               node:$value),
257                                     node:$addr),
258                     [{ return storeLoadCanUseBlockBinary(N, 1); }]>;
259 }
260 defm block_and : block_op<and>;
261 defm block_or  : block_op<or>;
262 defm block_xor : block_op<xor>;
263
264 // Insertions.
265 def inserti8 : PatFrag<(ops node:$src1, node:$src2),
266                        (or (and node:$src1, -256), node:$src2)>;
267 def insertll : PatFrag<(ops node:$src1, node:$src2),
268                        (or (and node:$src1, 0xffffffffffff0000), node:$src2)>;
269 def insertlh : PatFrag<(ops node:$src1, node:$src2),
270                        (or (and node:$src1, 0xffffffff0000ffff), node:$src2)>;
271 def inserthl : PatFrag<(ops node:$src1, node:$src2),
272                        (or (and node:$src1, 0xffff0000ffffffff), node:$src2)>;
273 def inserthh : PatFrag<(ops node:$src1, node:$src2),
274                        (or (and node:$src1, 0x0000ffffffffffff), node:$src2)>;
275 def insertlf : PatFrag<(ops node:$src1, node:$src2),
276                        (or (and node:$src1, 0xffffffff00000000), node:$src2)>;
277 def inserthf : PatFrag<(ops node:$src1, node:$src2),
278                        (or (and node:$src1, 0x00000000ffffffff), node:$src2)>;
279
280 // ORs that can be treated as insertions.
281 def or_as_inserti8 : PatFrag<(ops node:$src1, node:$src2),
282                              (or node:$src1, node:$src2), [{
283   unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
284   return CurDAG->MaskedValueIsZero(N->getOperand(0),
285                                    APInt::getLowBitsSet(BitWidth, 8));
286 }]>;
287
288 // ORs that can be treated as reversed insertions.
289 def or_as_revinserti8 : PatFrag<(ops node:$src1, node:$src2),
290                                 (or node:$src1, node:$src2), [{
291   unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
292   return CurDAG->MaskedValueIsZero(N->getOperand(1),
293                                    APInt::getLowBitsSet(BitWidth, 8));
294 }]>;
295
296 // Integer absolute, matching the canonical form generated by DAGCombiner.
297 def z_iabs32 : PatFrag<(ops node:$src),
298                        (xor (add node:$src, (sra node:$src, (i32 31))),
299                             (sra node:$src, (i32 31)))>;
300 def z_iabs64 : PatFrag<(ops node:$src),
301                        (xor (add node:$src, (sra node:$src, (i32 63))),
302                             (sra node:$src, (i32 63)))>;
303 def z_inegabs32 : PatFrag<(ops node:$src), (ineg (z_iabs32 node:$src))>;
304 def z_inegabs64 : PatFrag<(ops node:$src), (ineg (z_iabs64 node:$src))>;
305
306 // Fused multiply-add and multiply-subtract, but with the order of the
307 // operands matching SystemZ's MA and MS instructions.
308 def z_fma : PatFrag<(ops node:$src1, node:$src2, node:$src3),
309                     (fma node:$src2, node:$src3, node:$src1)>;
310 def z_fms : PatFrag<(ops node:$src1, node:$src2, node:$src3),
311                     (fma node:$src2, node:$src3, (fneg node:$src1))>;
312
313 // Floating-point negative absolute.
314 def fnabs : PatFrag<(ops node:$ptr), (fneg (fabs node:$ptr))>;
315
316 // Create a unary operator that loads from memory and then performs
317 // the given operation on it.
318 class loadu<SDPatternOperator operator, SDPatternOperator load = load>
319   : PatFrag<(ops node:$addr), (operator (load node:$addr))>;
320
321 // Create a store operator that performs the given unary operation
322 // on the value before storing it.
323 class storeu<SDPatternOperator operator, SDPatternOperator store = store>
324   : PatFrag<(ops node:$value, node:$addr),
325             (store (operator node:$value), node:$addr)>;