LLVM: lib/Target/X86/X86InstrFoldTables.cpp Source File (original) (raw)
1
2
3
4
5
6
7
8
9
10
11
12
16#include
17#include
18
19using namespace llvm;
20
21
22
23
24
25#include "X86GenFoldTables.inc"
26
27
28
30 { X86::VANDNPDZ128rr, X86::VANDNPSZ128rmb, TB_BCAST_SS },
31 { X86::VANDNPDZ256rr, X86::VANDNPSZ256rmb, TB_BCAST_SS },
32 { X86::VANDNPDZrr, X86::VANDNPSZrmb, TB_BCAST_SS },
33 { X86::VANDNPSZ128rr, X86::VANDNPDZ128rmb, TB_BCAST_SD },
34 { X86::VANDNPSZ256rr, X86::VANDNPDZ256rmb, TB_BCAST_SD },
35 { X86::VANDNPSZrr, X86::VANDNPDZrmb, TB_BCAST_SD },
36 { X86::VANDPDZ128rr, X86::VANDPSZ128rmb, TB_BCAST_SS },
37 { X86::VANDPDZ256rr, X86::VANDPSZ256rmb, TB_BCAST_SS },
38 { X86::VANDPDZrr, X86::VANDPSZrmb, TB_BCAST_SS },
39 { X86::VANDPSZ128rr, X86::VANDPDZ128rmb, TB_BCAST_SD },
40 { X86::VANDPSZ256rr, X86::VANDPDZ256rmb, TB_BCAST_SD },
41 { X86::VANDPSZrr, X86::VANDPDZrmb, TB_BCAST_SD },
42 { X86::VORPDZ128rr, X86::VORPSZ128rmb, TB_BCAST_SS },
43 { X86::VORPDZ256rr, X86::VORPSZ256rmb, TB_BCAST_SS },
44 { X86::VORPDZrr, X86::VORPSZrmb, TB_BCAST_SS },
45 { X86::VORPSZ128rr, X86::VORPDZ128rmb, TB_BCAST_SD },
46 { X86::VORPSZ256rr, X86::VORPDZ256rmb, TB_BCAST_SD },
47 { X86::VORPSZrr, X86::VORPDZrmb, TB_BCAST_SD },
48 { X86::VPANDDZ128rr, X86::VPANDQZ128rmb, TB_BCAST_Q },
49 { X86::VPANDDZ256rr, X86::VPANDQZ256rmb, TB_BCAST_Q },
50 { X86::VPANDDZrr, X86::VPANDQZrmb, TB_BCAST_Q },
51 { X86::VPANDNDZ128rr, X86::VPANDNQZ128rmb, TB_BCAST_Q },
52 { X86::VPANDNDZ256rr, X86::VPANDNQZ256rmb, TB_BCAST_Q },
53 { X86::VPANDNDZrr, X86::VPANDNQZrmb, TB_BCAST_Q },
54 { X86::VPANDNQZ128rr, X86::VPANDNDZ128rmb, TB_BCAST_D },
55 { X86::VPANDNQZ256rr, X86::VPANDNDZ256rmb, TB_BCAST_D },
56 { X86::VPANDNQZrr, X86::VPANDNDZrmb, TB_BCAST_D },
57 { X86::VPANDQZ128rr, X86::VPANDDZ128rmb, TB_BCAST_D },
58 { X86::VPANDQZ256rr, X86::VPANDDZ256rmb, TB_BCAST_D },
59 { X86::VPANDQZrr, X86::VPANDDZrmb, TB_BCAST_D },
60 { X86::VPORDZ128rr, X86::VPORQZ128rmb, TB_BCAST_Q },
61 { X86::VPORDZ256rr, X86::VPORQZ256rmb, TB_BCAST_Q },
62 { X86::VPORDZrr, X86::VPORQZrmb, TB_BCAST_Q },
63 { X86::VPORQZ128rr, X86::VPORDZ128rmb, TB_BCAST_D },
64 { X86::VPORQZ256rr, X86::VPORDZ256rmb, TB_BCAST_D },
65 { X86::VPORQZrr, X86::VPORDZrmb, TB_BCAST_D },
66 { X86::VPXORDZ128rr, X86::VPXORQZ128rmb, TB_BCAST_Q },
67 { X86::VPXORDZ256rr, X86::VPXORQZ256rmb, TB_BCAST_Q },
68 { X86::VPXORDZrr, X86::VPXORQZrmb, TB_BCAST_Q },
69 { X86::VPXORQZ128rr, X86::VPXORDZ128rmb, TB_BCAST_D },
70 { X86::VPXORQZ256rr, X86::VPXORDZ256rmb, TB_BCAST_D },
71 { X86::VPXORQZrr, X86::VPXORDZrmb, TB_BCAST_D },
72 { X86::VXORPDZ128rr, X86::VXORPSZ128rmb, TB_BCAST_SS },
73 { X86::VXORPDZ256rr, X86::VXORPSZ256rmb, TB_BCAST_SS },
74 { X86::VXORPDZrr, X86::VXORPSZrmb, TB_BCAST_SS },
75 { X86::VXORPSZ128rr, X86::VXORPDZ128rmb, TB_BCAST_SD },
76 { X86::VXORPSZ256rr, X86::VXORPDZ256rmb, TB_BCAST_SD },
77 { X86::VXORPSZrr, X86::VXORPDZrmb, TB_BCAST_SD },
78};
79
81 { X86::VPTERNLOGDZ128rri, X86::VPTERNLOGQZ128rmbi, TB_BCAST_Q },
82 { X86::VPTERNLOGDZ256rri, X86::VPTERNLOGQZ256rmbi, TB_BCAST_Q },
83 { X86::VPTERNLOGDZrri, X86::VPTERNLOGQZrmbi, TB_BCAST_Q },
84 { X86::VPTERNLOGQZ128rri, X86::VPTERNLOGDZ128rmbi, TB_BCAST_D },
85 { X86::VPTERNLOGQZ256rri, X86::VPTERNLOGDZ256rmbi, TB_BCAST_D },
86 { X86::VPTERNLOGQZrri, X86::VPTERNLOGDZrmbi, TB_BCAST_D },
87};
88
91#ifndef NDEBUG
92#define CHECK_SORTED_UNIQUE(TABLE) \
93 assert(llvm::is_sorted(TABLE) && #TABLE " is not sorted"); \
94 assert(std::adjacent_find(std::begin(Table), std::end(Table)) == \
95 std::end(Table) && \
96 #TABLE " is not unique");
97
98
99 static std::atomic FoldTablesChecked(false);
100 if (!FoldTablesChecked.load(std::memory_order_relaxed)) {
113 FoldTablesChecked.store(true, std::memory_order_relaxed);
114 }
115#endif
116
118 if (Data != Table.end() && Data->KeyOp == RegOp &&
120 return Data;
121 return nullptr;
122}
123
126}
127
130 if (OpNum == 0)
131 FoldTable = ArrayRef(Table0);
132 else if (OpNum == 1)
133 FoldTable = ArrayRef(Table1);
134 else if (OpNum == 2)
135 FoldTable = ArrayRef(Table2);
136 else if (OpNum == 3)
137 FoldTable = ArrayRef(Table3);
138 else if (OpNum == 4)
139 FoldTable = ArrayRef(Table4);
140 else
141 return nullptr;
142
144}
145
147 unsigned OpNum) {
149 if (OpNum == 1)
150 FoldTable = ArrayRef(BroadcastTable1);
151 else if (OpNum == 2)
152 FoldTable = ArrayRef(BroadcastTable2);
153 else if (OpNum == 3)
154 FoldTable = ArrayRef(BroadcastTable3);
155 else if (OpNum == 4)
156 FoldTable = ArrayRef(BroadcastTable4);
157 else
158 return nullptr;
159
161}
162
163namespace {
164
165
166
167struct X86MemUnfoldTable {
168
169 std::vector Table;
170
171 X86MemUnfoldTable() {
173
175
177
179
181
183
185
187
189
191
193
195
196
198
200
202
204
206
208
210
212
213
215
216
217 assert(std::adjacent_find(Table.begin(), Table.end()) == Table.end() &&
218 "Memory unfolding table is not unique!");
219 }
220
222
224 Table.push_back({Entry.DstOp, Entry.KeyOp,
225 static_cast<uint16_t>(Entry.Flags | ExtraFlags)});
226 }
227};
228}
229
231 static X86MemUnfoldTable MemUnfoldTable;
232 auto &Table = MemUnfoldTable.Table;
234 if (I != Table.end() && I->KeyOp == MemOp)
235 return &*I;
236 return nullptr;
237}
238
239namespace {
240
241
242
243struct X86BroadcastFoldTable {
244
245 std::vector Table;
246
247 X86BroadcastFoldTable() {
248
250 unsigned RegOp = Reg2Bcst.KeyOp;
251 unsigned BcstOp = Reg2Bcst.DstOp;
253 unsigned MemOp = Reg2Mem->DstOp;
256 Table.push_back({MemOp, BcstOp, Flags});
257 }
258 }
260 unsigned RegOp = Reg2Bcst.KeyOp;
261 unsigned BcstOp = Reg2Bcst.DstOp;
263 unsigned MemOp = Reg2Mem->DstOp;
266 Table.push_back({MemOp, BcstOp, Flags});
267 }
268 }
269
271 unsigned RegOp = Reg2Bcst.KeyOp;
272 unsigned BcstOp = Reg2Bcst.DstOp;
274 unsigned MemOp = Reg2Mem->DstOp;
277 Table.push_back({MemOp, BcstOp, Flags});
278 }
279 }
281 unsigned RegOp = Reg2Bcst.KeyOp;
282 unsigned BcstOp = Reg2Bcst.DstOp;
284 unsigned MemOp = Reg2Mem->DstOp;
287 Table.push_back({MemOp, BcstOp, Flags});
288 }
289 }
290
292 unsigned RegOp = Reg2Bcst.KeyOp;
293 unsigned BcstOp = Reg2Bcst.DstOp;
295 unsigned MemOp = Reg2Mem->DstOp;
298 Table.push_back({MemOp, BcstOp, Flags});
299 }
300 }
301
302
304 }
305};
306}
307
309 unsigned BroadcastBits) {
313 return BroadcastBits == 16;
316 return BroadcastBits == 32;
319 return BroadcastBits == 64;
320 }
321 return false;
322}
323
326 static X86BroadcastFoldTable BroadcastFoldTable;
327 auto &Table = BroadcastFoldTable.Table;
329 I != Table.end() && I->KeyOp == MemOp; ++I) {
331 return &*I;
332 }
333 return nullptr;
334}
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
static const X86FoldTableEntry BroadcastSizeTable2[]
#define CHECK_SORTED_UNIQUE(TABLE)
static const X86FoldTableEntry BroadcastSizeTable3[]
static const X86FoldTableEntry * lookupFoldTableImpl(ArrayRef< X86FoldTableEntry > Table, unsigned RegOp)
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
This is an optimization pass for GlobalISel generic memory operations.
const X86FoldTableEntry * lookupBroadcastFoldTableBySize(unsigned MemOp, unsigned BroadcastBits)
const X86FoldTableEntry * lookupBroadcastFoldTable(unsigned RegOp, unsigned OpNum)
const X86FoldTableEntry * lookupTwoAddrFoldTable(unsigned RegOp)
auto lower_bound(R &&Range, T &&Value)
Provide wrappers to std::lower_bound which take ranges instead of having to pass begin/end explicitly...
const X86FoldTableEntry * lookupUnfoldTable(unsigned MemOp)
bool matchBroadcastSize(const X86FoldTableEntry &Entry, unsigned BroadcastBits)
void array_pod_sort(IteratorTy Start, IteratorTy End)
array_pod_sort - This sorts an array with the specified start and end extent.
const X86FoldTableEntry * lookupFoldTable(unsigned RegOp, unsigned OpNum)