[SystemZ] Add TM and TMY
[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_ZICmp               : SDTypeProfile<0, 3,
19                                             [SDTCisSameAs<0, 1>,
20                                              SDTCisVT<2, i32>]>;
21 def SDT_ZBRCCMask           : SDTypeProfile<0, 3,
22                                             [SDTCisVT<0, i8>,
23                                              SDTCisVT<1, i8>,
24                                              SDTCisVT<2, OtherVT>]>;
25 def SDT_ZSelectCCMask       : SDTypeProfile<1, 4,
26                                             [SDTCisSameAs<0, 1>,
27                                              SDTCisSameAs<1, 2>,
28                                              SDTCisVT<3, i8>,
29                                              SDTCisVT<4, i8>]>;
30 def SDT_ZWrapPtr            : SDTypeProfile<1, 1,
31                                             [SDTCisSameAs<0, 1>,
32                                              SDTCisPtrTy<0>]>;
33 def SDT_ZAdjDynAlloc        : SDTypeProfile<1, 0, [SDTCisVT<0, i64>]>;
34 def SDT_ZExtractAccess      : SDTypeProfile<1, 1,
35                                             [SDTCisVT<0, i32>,
36                                              SDTCisVT<1, i8>]>;
37 def SDT_ZGR128Binary32      : SDTypeProfile<1, 2,
38                                             [SDTCisVT<0, untyped>,
39                                              SDTCisVT<1, untyped>,
40                                              SDTCisVT<2, i32>]>;
41 def SDT_ZGR128Binary64      : SDTypeProfile<1, 2,
42                                             [SDTCisVT<0, untyped>,
43                                              SDTCisVT<1, untyped>,
44                                              SDTCisVT<2, i64>]>;
45 def SDT_ZAtomicLoadBinaryW  : SDTypeProfile<1, 5,
46                                             [SDTCisVT<0, i32>,
47                                              SDTCisPtrTy<1>,
48                                              SDTCisVT<2, i32>,
49                                              SDTCisVT<3, i32>,
50                                              SDTCisVT<4, i32>,
51                                              SDTCisVT<5, i32>]>;
52 def SDT_ZAtomicCmpSwapW     : SDTypeProfile<1, 6,
53                                             [SDTCisVT<0, i32>,
54                                              SDTCisPtrTy<1>,
55                                              SDTCisVT<2, i32>,
56                                              SDTCisVT<3, i32>,
57                                              SDTCisVT<4, i32>,
58                                              SDTCisVT<5, i32>,
59                                              SDTCisVT<6, i32>]>;
60 def SDT_ZMemMemLength       : SDTypeProfile<0, 3,
61                                             [SDTCisPtrTy<0>,
62                                              SDTCisPtrTy<1>,
63                                              SDTCisVT<2, i64>]>;
64 def SDT_ZMemMemLoop         : SDTypeProfile<0, 4,
65                                             [SDTCisPtrTy<0>,
66                                              SDTCisPtrTy<1>,
67                                              SDTCisVT<2, i64>,
68                                              SDTCisVT<3, i64>]>;
69 def SDT_ZString             : SDTypeProfile<1, 3,
70                                             [SDTCisPtrTy<0>,
71                                              SDTCisPtrTy<1>,
72                                              SDTCisPtrTy<2>,
73                                              SDTCisVT<3, i32>]>;
74 def SDT_ZI32Intrinsic       : SDTypeProfile<1, 0, [SDTCisVT<0, i32>]>;
75 def SDT_ZPrefetch           : SDTypeProfile<0, 2,
76                                             [SDTCisVT<0, i8>,
77                                              SDTCisPtrTy<1>]>;
78
79 //===----------------------------------------------------------------------===//
80 // Node definitions
81 //===----------------------------------------------------------------------===//
82
83 // These are target-independent nodes, but have target-specific formats.
84 def callseq_start       : SDNode<"ISD::CALLSEQ_START", SDT_CallSeqStart,
85                                  [SDNPHasChain, SDNPSideEffect, SDNPOutGlue]>;
86 def callseq_end         : SDNode<"ISD::CALLSEQ_END",   SDT_CallSeqEnd,
87                                  [SDNPHasChain, SDNPSideEffect, SDNPOptInGlue,
88                                   SDNPOutGlue]>;
89
90 // Nodes for SystemZISD::*.  See SystemZISelLowering.h for more details.
91 def z_retflag           : SDNode<"SystemZISD::RET_FLAG", SDTNone,
92                                  [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
93 def z_call              : SDNode<"SystemZISD::CALL", SDT_ZCall,
94                                  [SDNPHasChain, SDNPOutGlue, SDNPOptInGlue,
95                                   SDNPVariadic]>;
96 def z_sibcall           : SDNode<"SystemZISD::SIBCALL", SDT_ZCall,
97                                  [SDNPHasChain, SDNPOutGlue, SDNPOptInGlue,
98                                   SDNPVariadic]>;
99 def z_pcrel_wrapper     : SDNode<"SystemZISD::PCREL_WRAPPER", SDT_ZWrapPtr, []>;
100 def z_icmp              : SDNode<"SystemZISD::ICMP", SDT_ZICmp, [SDNPOutGlue]>;
101 def z_fcmp              : SDNode<"SystemZISD::FCMP", SDT_ZCmp, [SDNPOutGlue]>;
102 def z_tm                : SDNode<"SystemZISD::TM", SDT_ZICmp, [SDNPOutGlue]>;
103 def z_br_ccmask         : SDNode<"SystemZISD::BR_CCMASK", SDT_ZBRCCMask,
104                                  [SDNPHasChain, SDNPInGlue]>;
105 def z_select_ccmask     : SDNode<"SystemZISD::SELECT_CCMASK", SDT_ZSelectCCMask,
106                                  [SDNPInGlue]>;
107 def z_adjdynalloc       : SDNode<"SystemZISD::ADJDYNALLOC", SDT_ZAdjDynAlloc>;
108 def z_extract_access    : SDNode<"SystemZISD::EXTRACT_ACCESS",
109                                  SDT_ZExtractAccess>;
110 def z_umul_lohi64       : SDNode<"SystemZISD::UMUL_LOHI64", SDT_ZGR128Binary64>;
111 def z_sdivrem32         : SDNode<"SystemZISD::SDIVREM32", SDT_ZGR128Binary32>;
112 def z_sdivrem64         : SDNode<"SystemZISD::SDIVREM64", SDT_ZGR128Binary64>;
113 def z_udivrem32         : SDNode<"SystemZISD::UDIVREM32", SDT_ZGR128Binary32>;
114 def z_udivrem64         : SDNode<"SystemZISD::UDIVREM64", SDT_ZGR128Binary64>;
115
116 class AtomicWOp<string name, SDTypeProfile profile = SDT_ZAtomicLoadBinaryW>
117   : SDNode<"SystemZISD::"##name, profile,
118            [SDNPHasChain, SDNPMayStore, SDNPMayLoad, SDNPMemOperand]>;
119
120 def z_atomic_swapw      : AtomicWOp<"ATOMIC_SWAPW">;
121 def z_atomic_loadw_add  : AtomicWOp<"ATOMIC_LOADW_ADD">;
122 def z_atomic_loadw_sub  : AtomicWOp<"ATOMIC_LOADW_SUB">;
123 def z_atomic_loadw_and  : AtomicWOp<"ATOMIC_LOADW_AND">;
124 def z_atomic_loadw_or   : AtomicWOp<"ATOMIC_LOADW_OR">;
125 def z_atomic_loadw_xor  : AtomicWOp<"ATOMIC_LOADW_XOR">;
126 def z_atomic_loadw_nand : AtomicWOp<"ATOMIC_LOADW_NAND">;
127 def z_atomic_loadw_min  : AtomicWOp<"ATOMIC_LOADW_MIN">;
128 def z_atomic_loadw_max  : AtomicWOp<"ATOMIC_LOADW_MAX">;
129 def z_atomic_loadw_umin : AtomicWOp<"ATOMIC_LOADW_UMIN">;
130 def z_atomic_loadw_umax : AtomicWOp<"ATOMIC_LOADW_UMAX">;
131 def z_atomic_cmp_swapw  : AtomicWOp<"ATOMIC_CMP_SWAPW", SDT_ZAtomicCmpSwapW>;
132
133 def z_mvc               : SDNode<"SystemZISD::MVC", SDT_ZMemMemLength,
134                                  [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
135 def z_mvc_loop          : SDNode<"SystemZISD::MVC_LOOP", SDT_ZMemMemLoop,
136                                  [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
137 def z_nc                : SDNode<"SystemZISD::NC", SDT_ZMemMemLength,
138                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
139 def z_nc_loop           : SDNode<"SystemZISD::NC_LOOP", SDT_ZMemMemLoop,
140                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
141 def z_oc                : SDNode<"SystemZISD::OC", SDT_ZMemMemLength,
142                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
143 def z_oc_loop           : SDNode<"SystemZISD::OC_LOOP", SDT_ZMemMemLoop,
144                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
145 def z_xc                : SDNode<"SystemZISD::XC", SDT_ZMemMemLength,
146                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
147 def z_xc_loop           : SDNode<"SystemZISD::XC_LOOP", SDT_ZMemMemLoop,
148                                   [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
149 def z_clc               : SDNode<"SystemZISD::CLC", SDT_ZMemMemLength,
150                                  [SDNPHasChain, SDNPOutGlue, SDNPMayLoad]>;
151 def z_clc_loop          : SDNode<"SystemZISD::CLC_LOOP", SDT_ZMemMemLoop,
152                                  [SDNPHasChain, SDNPOutGlue, SDNPMayLoad]>;
153 def z_strcmp            : SDNode<"SystemZISD::STRCMP", SDT_ZString,
154                                  [SDNPHasChain, SDNPOutGlue, SDNPMayLoad]>;
155 def z_stpcpy            : SDNode<"SystemZISD::STPCPY", SDT_ZString,
156                                  [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
157 def z_search_string     : SDNode<"SystemZISD::SEARCH_STRING", SDT_ZString,
158                                  [SDNPHasChain, SDNPOutGlue, SDNPMayLoad]>;
159 def z_ipm               : SDNode<"SystemZISD::IPM", SDT_ZI32Intrinsic,
160                                  [SDNPInGlue]>;
161 def z_prefetch          : SDNode<"SystemZISD::PREFETCH", SDT_ZPrefetch,
162                                  [SDNPHasChain, SDNPMayLoad, SDNPMayStore,
163                                   SDNPMemOperand]>;
164
165 //===----------------------------------------------------------------------===//
166 // Pattern fragments
167 //===----------------------------------------------------------------------===//
168
169 // Signed and unsigned comparisons.
170 def z_scmp : PatFrag<(ops node:$a, node:$b), (z_icmp node:$a, node:$b, imm), [{
171   unsigned Type = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
172   return Type != SystemZICMP::UnsignedOnly;
173 }]>;
174 def z_ucmp : PatFrag<(ops node:$a, node:$b), (z_icmp node:$a, node:$b, imm), [{
175   unsigned Type = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
176   return Type != SystemZICMP::SignedOnly;
177 }]>;
178
179 // Register- and memory-based TEST UNDER MASK.
180 def z_tm_reg : PatFrag<(ops node:$a, node:$b), (z_tm node:$a, node:$b, imm)>;
181 def z_tm_mem : PatFrag<(ops node:$a, node:$b), (z_tm node:$a, node:$b, 0)>;
182
183 // Register sign-extend operations.  Sub-32-bit values are represented as i32s.
184 def sext8  : PatFrag<(ops node:$src), (sext_inreg node:$src, i8)>;
185 def sext16 : PatFrag<(ops node:$src), (sext_inreg node:$src, i16)>;
186 def sext32 : PatFrag<(ops node:$src), (sext (i32 node:$src))>;
187
188 // Register zero-extend operations.  Sub-32-bit values are represented as i32s.
189 def zext8  : PatFrag<(ops node:$src), (and node:$src, 0xff)>;
190 def zext16 : PatFrag<(ops node:$src), (and node:$src, 0xffff)>;
191 def zext32 : PatFrag<(ops node:$src), (zext (i32 node:$src))>;
192
193 // Typed floating-point loads.
194 def loadf32 : PatFrag<(ops node:$src), (f32 (load node:$src))>;
195 def loadf64 : PatFrag<(ops node:$src), (f64 (load node:$src))>;
196
197 // Extending loads in which the extension type doesn't matter.
198 def anyextload : PatFrag<(ops node:$ptr), (unindexedload node:$ptr), [{
199   return cast<LoadSDNode>(N)->getExtensionType() != ISD::NON_EXTLOAD;
200 }]>;
201 def anyextloadi8 : PatFrag<(ops node:$ptr), (anyextload node:$ptr), [{
202   return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i8;
203 }]>;
204 def anyextloadi16 : PatFrag<(ops node:$ptr), (anyextload node:$ptr), [{
205   return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i16;
206 }]>;
207 def anyextloadi32 : PatFrag<(ops node:$ptr), (anyextload node:$ptr), [{
208   return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i32;
209 }]>;
210
211 // Aligned loads.
212 class AlignedLoad<SDPatternOperator load>
213   : PatFrag<(ops node:$addr), (load node:$addr), [{
214   LoadSDNode *Load = cast<LoadSDNode>(N);
215   return Load->getAlignment() >= Load->getMemoryVT().getStoreSize();
216 }]>;
217 def aligned_load        : AlignedLoad<load>;
218 def aligned_sextloadi16 : AlignedLoad<sextloadi16>;
219 def aligned_sextloadi32 : AlignedLoad<sextloadi32>;
220 def aligned_zextloadi16 : AlignedLoad<zextloadi16>;
221 def aligned_zextloadi32 : AlignedLoad<zextloadi32>;
222
223 // Aligned stores.
224 class AlignedStore<SDPatternOperator store>
225   : PatFrag<(ops node:$src, node:$addr), (store node:$src, node:$addr), [{
226   StoreSDNode *Store = cast<StoreSDNode>(N);
227   return Store->getAlignment() >= Store->getMemoryVT().getStoreSize();
228 }]>;
229 def aligned_store         : AlignedStore<store>;
230 def aligned_truncstorei16 : AlignedStore<truncstorei16>;
231 def aligned_truncstorei32 : AlignedStore<truncstorei32>;
232
233 // Non-volatile loads.  Used for instructions that might access the storage
234 // location multiple times.
235 class NonvolatileLoad<SDPatternOperator load>
236   : PatFrag<(ops node:$addr), (load node:$addr), [{
237   LoadSDNode *Load = cast<LoadSDNode>(N);
238   return !Load->isVolatile();
239 }]>;
240 def nonvolatile_load          : NonvolatileLoad<load>;
241 def nonvolatile_anyextloadi8  : NonvolatileLoad<anyextloadi8>;
242 def nonvolatile_anyextloadi16 : NonvolatileLoad<anyextloadi16>;
243 def nonvolatile_anyextloadi32 : NonvolatileLoad<anyextloadi32>;
244
245 // Non-volatile stores.
246 class NonvolatileStore<SDPatternOperator store>
247   : PatFrag<(ops node:$src, node:$addr), (store node:$src, node:$addr), [{
248   StoreSDNode *Store = cast<StoreSDNode>(N);
249   return !Store->isVolatile();
250 }]>;
251 def nonvolatile_store         : NonvolatileStore<store>;
252 def nonvolatile_truncstorei8  : NonvolatileStore<truncstorei8>;
253 def nonvolatile_truncstorei16 : NonvolatileStore<truncstorei16>;
254 def nonvolatile_truncstorei32 : NonvolatileStore<truncstorei32>;
255
256 // A store of a load that can be implemented using MVC.
257 def mvc_store : PatFrag<(ops node:$value, node:$addr),
258                         (unindexedstore node:$value, node:$addr),
259                         [{ return storeLoadCanUseMVC(N); }]>;
260
261 // Binary read-modify-write operations on memory in which the other
262 // operand is also memory and for which block operations like NC can
263 // be used.  There are two patterns for each operator, depending on
264 // which operand contains the "other" load.
265 multiclass block_op<SDPatternOperator operator> {
266   def "1" : PatFrag<(ops node:$value, node:$addr),
267                     (unindexedstore (operator node:$value,
268                                               (unindexedload node:$addr)),
269                                     node:$addr),
270                     [{ return storeLoadCanUseBlockBinary(N, 0); }]>;
271   def "2" : PatFrag<(ops node:$value, node:$addr),
272                     (unindexedstore (operator (unindexedload node:$addr),
273                                               node:$value),
274                                     node:$addr),
275                     [{ return storeLoadCanUseBlockBinary(N, 1); }]>;
276 }
277 defm block_and : block_op<and>;
278 defm block_or  : block_op<or>;
279 defm block_xor : block_op<xor>;
280
281 // Insertions.
282 def inserti8 : PatFrag<(ops node:$src1, node:$src2),
283                        (or (and node:$src1, -256), node:$src2)>;
284 def insertll : PatFrag<(ops node:$src1, node:$src2),
285                        (or (and node:$src1, 0xffffffffffff0000), node:$src2)>;
286 def insertlh : PatFrag<(ops node:$src1, node:$src2),
287                        (or (and node:$src1, 0xffffffff0000ffff), node:$src2)>;
288 def inserthl : PatFrag<(ops node:$src1, node:$src2),
289                        (or (and node:$src1, 0xffff0000ffffffff), node:$src2)>;
290 def inserthh : PatFrag<(ops node:$src1, node:$src2),
291                        (or (and node:$src1, 0x0000ffffffffffff), node:$src2)>;
292 def insertlf : PatFrag<(ops node:$src1, node:$src2),
293                        (or (and node:$src1, 0xffffffff00000000), node:$src2)>;
294 def inserthf : PatFrag<(ops node:$src1, node:$src2),
295                        (or (and node:$src1, 0x00000000ffffffff), node:$src2)>;
296
297 // ORs that can be treated as insertions.
298 def or_as_inserti8 : PatFrag<(ops node:$src1, node:$src2),
299                              (or node:$src1, node:$src2), [{
300   unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
301   return CurDAG->MaskedValueIsZero(N->getOperand(0),
302                                    APInt::getLowBitsSet(BitWidth, 8));
303 }]>;
304
305 // ORs that can be treated as reversed insertions.
306 def or_as_revinserti8 : PatFrag<(ops node:$src1, node:$src2),
307                                 (or node:$src1, node:$src2), [{
308   unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
309   return CurDAG->MaskedValueIsZero(N->getOperand(1),
310                                    APInt::getLowBitsSet(BitWidth, 8));
311 }]>;
312
313 // Integer absolute, matching the canonical form generated by DAGCombiner.
314 def z_iabs32 : PatFrag<(ops node:$src),
315                        (xor (add node:$src, (sra node:$src, (i32 31))),
316                             (sra node:$src, (i32 31)))>;
317 def z_iabs64 : PatFrag<(ops node:$src),
318                        (xor (add node:$src, (sra node:$src, (i32 63))),
319                             (sra node:$src, (i32 63)))>;
320 def z_inegabs32 : PatFrag<(ops node:$src), (ineg (z_iabs32 node:$src))>;
321 def z_inegabs64 : PatFrag<(ops node:$src), (ineg (z_iabs64 node:$src))>;
322
323 // Fused multiply-add and multiply-subtract, but with the order of the
324 // operands matching SystemZ's MA and MS instructions.
325 def z_fma : PatFrag<(ops node:$src1, node:$src2, node:$src3),
326                     (fma node:$src2, node:$src3, node:$src1)>;
327 def z_fms : PatFrag<(ops node:$src1, node:$src2, node:$src3),
328                     (fma node:$src2, node:$src3, (fneg node:$src1))>;
329
330 // Floating-point negative absolute.
331 def fnabs : PatFrag<(ops node:$ptr), (fneg (fabs node:$ptr))>;
332
333 // Create a unary operator that loads from memory and then performs
334 // the given operation on it.
335 class loadu<SDPatternOperator operator, SDPatternOperator load = load>
336   : PatFrag<(ops node:$addr), (operator (load node:$addr))>;
337
338 // Create a store operator that performs the given unary operation
339 // on the value before storing it.
340 class storeu<SDPatternOperator operator, SDPatternOperator store = store>
341   : PatFrag<(ops node:$value, node:$addr),
342             (store (operator node:$value), node:$addr)>;