[SystemZ] Use CLC and IPM to implement memcmp
[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, i32>]>;
61 def SDT_ZI32Intrinsic       : SDTypeProfile<1, 0, [SDTCisVT<0, i32>]>;
62
63 //===----------------------------------------------------------------------===//
64 // Node definitions
65 //===----------------------------------------------------------------------===//
66
67 // These are target-independent nodes, but have target-specific formats.
68 def callseq_start       : SDNode<"ISD::CALLSEQ_START", SDT_CallSeqStart,
69                                  [SDNPHasChain, SDNPSideEffect, SDNPOutGlue]>;
70 def callseq_end         : SDNode<"ISD::CALLSEQ_END",   SDT_CallSeqEnd,
71                                  [SDNPHasChain, SDNPSideEffect, SDNPOptInGlue,
72                                   SDNPOutGlue]>;
73
74 // Nodes for SystemZISD::*.  See SystemZISelLowering.h for more details.
75 def z_retflag           : SDNode<"SystemZISD::RET_FLAG", SDTNone,
76                                  [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
77 def z_call              : SDNode<"SystemZISD::CALL", SDT_ZCall,
78                                  [SDNPHasChain, SDNPOutGlue, SDNPOptInGlue,
79                                   SDNPVariadic]>;
80 def z_pcrel_wrapper     : SDNode<"SystemZISD::PCREL_WRAPPER", SDT_ZWrapPtr, []>;
81 def z_cmp               : SDNode<"SystemZISD::CMP", SDT_ZCmp, [SDNPOutGlue]>;
82 def z_ucmp              : SDNode<"SystemZISD::UCMP", SDT_ZCmp, [SDNPOutGlue]>;
83 def z_br_ccmask         : SDNode<"SystemZISD::BR_CCMASK", SDT_ZBRCCMask,
84                                  [SDNPHasChain, SDNPInGlue]>;
85 def z_select_ccmask     : SDNode<"SystemZISD::SELECT_CCMASK", SDT_ZSelectCCMask,
86                                  [SDNPInGlue]>;
87 def z_adjdynalloc       : SDNode<"SystemZISD::ADJDYNALLOC", SDT_ZAdjDynAlloc>;
88 def z_extract_access    : SDNode<"SystemZISD::EXTRACT_ACCESS",
89                                  SDT_ZExtractAccess>;
90 def z_umul_lohi64       : SDNode<"SystemZISD::UMUL_LOHI64", SDT_ZGR128Binary64>;
91 def z_sdivrem32         : SDNode<"SystemZISD::SDIVREM32", SDT_ZGR128Binary32>;
92 def z_sdivrem64         : SDNode<"SystemZISD::SDIVREM64", SDT_ZGR128Binary64>;
93 def z_udivrem32         : SDNode<"SystemZISD::UDIVREM32", SDT_ZGR128Binary32>;
94 def z_udivrem64         : SDNode<"SystemZISD::UDIVREM64", SDT_ZGR128Binary64>;
95
96 class AtomicWOp<string name, SDTypeProfile profile = SDT_ZAtomicLoadBinaryW>
97   : SDNode<"SystemZISD::"##name, profile,
98            [SDNPHasChain, SDNPMayStore, SDNPMayLoad, SDNPMemOperand]>;
99
100 def z_atomic_swapw      : AtomicWOp<"ATOMIC_SWAPW">;
101 def z_atomic_loadw_add  : AtomicWOp<"ATOMIC_LOADW_ADD">;
102 def z_atomic_loadw_sub  : AtomicWOp<"ATOMIC_LOADW_SUB">;
103 def z_atomic_loadw_and  : AtomicWOp<"ATOMIC_LOADW_AND">;
104 def z_atomic_loadw_or   : AtomicWOp<"ATOMIC_LOADW_OR">;
105 def z_atomic_loadw_xor  : AtomicWOp<"ATOMIC_LOADW_XOR">;
106 def z_atomic_loadw_nand : AtomicWOp<"ATOMIC_LOADW_NAND">;
107 def z_atomic_loadw_min  : AtomicWOp<"ATOMIC_LOADW_MIN">;
108 def z_atomic_loadw_max  : AtomicWOp<"ATOMIC_LOADW_MAX">;
109 def z_atomic_loadw_umin : AtomicWOp<"ATOMIC_LOADW_UMIN">;
110 def z_atomic_loadw_umax : AtomicWOp<"ATOMIC_LOADW_UMAX">;
111 def z_atomic_cmp_swapw  : AtomicWOp<"ATOMIC_CMP_SWAPW", SDT_ZAtomicCmpSwapW>;
112
113 def z_mvc               : SDNode<"SystemZISD::MVC", SDT_ZMemMemLength,
114                                  [SDNPHasChain, SDNPMayStore, SDNPMayLoad]>;
115 def z_clc               : SDNode<"SystemZISD::CLC", SDT_ZMemMemLength,
116                                  [SDNPHasChain, SDNPOutGlue, SDNPMayLoad]>;
117 def z_ipm               : SDNode<"SystemZISD::IPM", SDT_ZI32Intrinsic,
118                                  [SDNPInGlue]>;
119
120 //===----------------------------------------------------------------------===//
121 // Pattern fragments
122 //===----------------------------------------------------------------------===//
123
124 // Register sign-extend operations.  Sub-32-bit values are represented as i32s.
125 def sext8  : PatFrag<(ops node:$src), (sext_inreg node:$src, i8)>;
126 def sext16 : PatFrag<(ops node:$src), (sext_inreg node:$src, i16)>;
127 def sext32 : PatFrag<(ops node:$src), (sext (i32 node:$src))>;
128
129 // Register zero-extend operations.  Sub-32-bit values are represented as i32s.
130 def zext8  : PatFrag<(ops node:$src), (and node:$src, 0xff)>;
131 def zext16 : PatFrag<(ops node:$src), (and node:$src, 0xffff)>;
132 def zext32 : PatFrag<(ops node:$src), (zext (i32 node:$src))>;
133
134 // Typed floating-point loads.
135 def loadf32 : PatFrag<(ops node:$src), (f32 (load node:$src))>;
136 def loadf64 : PatFrag<(ops node:$src), (f64 (load node:$src))>;
137
138 // Extending loads in which the extension type doesn't matter.
139 def anyextload : PatFrag<(ops node:$ptr), (unindexedload node:$ptr), [{
140   return cast<LoadSDNode>(N)->getExtensionType() != ISD::NON_EXTLOAD;
141 }]>;
142 def anyextloadi8 : PatFrag<(ops node:$ptr), (anyextload node:$ptr), [{
143   return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i8;
144 }]>;
145 def anyextloadi16 : PatFrag<(ops node:$ptr), (anyextload node:$ptr), [{
146   return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i16;
147 }]>;
148 def anyextloadi32 : PatFrag<(ops node:$ptr), (anyextload node:$ptr), [{
149   return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i32;
150 }]>;
151
152 // Aligned loads.
153 class AlignedLoad<SDPatternOperator load>
154   : PatFrag<(ops node:$addr), (load node:$addr), [{
155   LoadSDNode *Load = cast<LoadSDNode>(N);
156   return Load->getAlignment() >= Load->getMemoryVT().getStoreSize();
157 }]>;
158 def aligned_load        : AlignedLoad<load>;
159 def aligned_sextloadi16 : AlignedLoad<sextloadi16>;
160 def aligned_sextloadi32 : AlignedLoad<sextloadi32>;
161 def aligned_zextloadi16 : AlignedLoad<zextloadi16>;
162 def aligned_zextloadi32 : AlignedLoad<zextloadi32>;
163
164 // Aligned stores.
165 class AlignedStore<SDPatternOperator store>
166   : PatFrag<(ops node:$src, node:$addr), (store node:$src, node:$addr), [{
167   StoreSDNode *Store = cast<StoreSDNode>(N);
168   return Store->getAlignment() >= Store->getMemoryVT().getStoreSize();
169 }]>;
170 def aligned_store         : AlignedStore<store>;
171 def aligned_truncstorei16 : AlignedStore<truncstorei16>;
172 def aligned_truncstorei32 : AlignedStore<truncstorei32>;
173
174 // Non-volatile loads.  Used for instructions that might access the storage
175 // location multiple times.
176 class NonvolatileLoad<SDPatternOperator load>
177   : PatFrag<(ops node:$addr), (load node:$addr), [{
178   LoadSDNode *Load = cast<LoadSDNode>(N);
179   return !Load->isVolatile();
180 }]>;
181 def nonvolatile_load          : NonvolatileLoad<load>;
182 def nonvolatile_anyextloadi8  : NonvolatileLoad<anyextloadi8>;
183 def nonvolatile_anyextloadi16 : NonvolatileLoad<anyextloadi16>;
184 def nonvolatile_anyextloadi32 : NonvolatileLoad<anyextloadi32>;
185
186 // Non-volatile stores.
187 class NonvolatileStore<SDPatternOperator store>
188   : PatFrag<(ops node:$src, node:$addr), (store node:$src, node:$addr), [{
189   StoreSDNode *Store = cast<StoreSDNode>(N);
190   return !Store->isVolatile();
191 }]>;
192 def nonvolatile_store         : NonvolatileStore<store>;
193 def nonvolatile_truncstorei8  : NonvolatileStore<truncstorei8>;
194 def nonvolatile_truncstorei16 : NonvolatileStore<truncstorei16>;
195 def nonvolatile_truncstorei32 : NonvolatileStore<truncstorei32>;
196
197 // Insertions.
198 def inserti8 : PatFrag<(ops node:$src1, node:$src2),
199                        (or (and node:$src1, -256), node:$src2)>;
200 def insertll : PatFrag<(ops node:$src1, node:$src2),
201                        (or (and node:$src1, 0xffffffffffff0000), node:$src2)>;
202 def insertlh : PatFrag<(ops node:$src1, node:$src2),
203                        (or (and node:$src1, 0xffffffff0000ffff), node:$src2)>;
204 def inserthl : PatFrag<(ops node:$src1, node:$src2),
205                        (or (and node:$src1, 0xffff0000ffffffff), node:$src2)>;
206 def inserthh : PatFrag<(ops node:$src1, node:$src2),
207                        (or (and node:$src1, 0x0000ffffffffffff), node:$src2)>;
208 def insertlf : PatFrag<(ops node:$src1, node:$src2),
209                        (or (and node:$src1, 0xffffffff00000000), node:$src2)>;
210 def inserthf : PatFrag<(ops node:$src1, node:$src2),
211                        (or (and node:$src1, 0x00000000ffffffff), node:$src2)>;
212
213 // ORs that can be treated as insertions.
214 def or_as_inserti8 : PatFrag<(ops node:$src1, node:$src2),
215                              (or node:$src1, node:$src2), [{
216   unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
217   return CurDAG->MaskedValueIsZero(N->getOperand(0),
218                                    APInt::getLowBitsSet(BitWidth, 8));
219 }]>;
220
221 // ORs that can be treated as reversed insertions.
222 def or_as_revinserti8 : PatFrag<(ops node:$src1, node:$src2),
223                                 (or node:$src1, node:$src2), [{
224   unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
225   return CurDAG->MaskedValueIsZero(N->getOperand(1),
226                                    APInt::getLowBitsSet(BitWidth, 8));
227 }]>;
228
229 // Fused multiply-add and multiply-subtract, but with the order of the
230 // operands matching SystemZ's MA and MS instructions.
231 def z_fma : PatFrag<(ops node:$src1, node:$src2, node:$src3),
232                     (fma node:$src2, node:$src3, node:$src1)>;
233 def z_fms : PatFrag<(ops node:$src1, node:$src2, node:$src3),
234                     (fma node:$src2, node:$src3, (fneg node:$src1))>;
235
236 // Floating-point negative absolute.
237 def fnabs : PatFrag<(ops node:$ptr), (fneg (fabs node:$ptr))>;
238
239 // Create a unary operator that loads from memory and then performs
240 // the given operation on it.
241 class loadu<SDPatternOperator operator, SDPatternOperator load = load>
242   : PatFrag<(ops node:$addr), (operator (load node:$addr))>;
243
244 // Create a store operator that performs the given unary operation
245 // on the value before storing it.
246 class storeu<SDPatternOperator operator, SDPatternOperator store = store>
247   : PatFrag<(ops node:$value, node:$addr),
248             (store (operator node:$value), node:$addr)>;