xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/AArch64Combine.td (revision 1db9f3b21e39176dd5b67cf8ac378633b172463e)
1//=- AArch64Combine.td - Define AArch64 Combine Rules ---------*-tablegen -*-=//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9//
10//===----------------------------------------------------------------------===//
11
12include "llvm/Target/GlobalISel/Combine.td"
13
14def fconstant_to_constant : GICombineRule<
15  (defs root:$root),
16  (match (wip_match_opcode G_FCONSTANT):$root,
17         [{ return matchFConstantToConstant(*${root}, MRI); }]),
18  (apply [{ applyFConstantToConstant(*${root}); }])>;
19
20def icmp_redundant_trunc_matchdata : GIDefMatchData<"Register">;
21def icmp_redundant_trunc : GICombineRule<
22  (defs root:$root, icmp_redundant_trunc_matchdata:$matchinfo),
23  (match (wip_match_opcode G_ICMP):$root,
24         [{ return matchICmpRedundantTrunc(*${root}, MRI, Helper.getKnownBits(), ${matchinfo}); }]),
25  (apply [{ applyICmpRedundantTrunc(*${root}, MRI, B, Observer, ${matchinfo}); }])>;
26
27// AArch64-specific offset folding for G_GLOBAL_VALUE.
28def fold_global_offset_matchdata : GIDefMatchData<"std::pair<uint64_t, uint64_t>">;
29def fold_global_offset : GICombineRule<
30  (defs root:$root, fold_global_offset_matchdata:$matchinfo),
31  (match (wip_match_opcode G_GLOBAL_VALUE):$root,
32          [{ return matchFoldGlobalOffset(*${root}, MRI, ${matchinfo}); }]),
33  (apply [{ applyFoldGlobalOffset(*${root}, MRI, B, Observer, ${matchinfo});}])
34>;
35
36// Boolean: 0 = G_ZEXT, 1 = G_SEXT
37def ext_addv_to_udot_addv_matchinfo : GIDefMatchData<"std::tuple<Register, Register, bool>">;
38let Predicates = [HasDotProd] in {
39def ext_addv_to_udot_addv : GICombineRule<
40  (defs root:$root, ext_addv_to_udot_addv_matchinfo:$matchinfo),
41  (match (wip_match_opcode G_VECREDUCE_ADD):$root,
42         [{ return matchExtAddvToUdotAddv(*${root}, MRI, STI, ${matchinfo}); }]),
43  (apply [{ applyExtAddvToUdotAddv(*${root}, MRI, B, Observer, STI, ${matchinfo}); }])
44>;
45}
46
47def AArch64PreLegalizerCombiner: GICombiner<
48  "AArch64PreLegalizerCombinerImpl", [all_combines,
49                                      fconstant_to_constant,
50                                      icmp_redundant_trunc,
51                                      fold_global_offset,
52                                      shuffle_to_extract,
53                                      ext_addv_to_udot_addv]> {
54  let CombineAllMethodName = "tryCombineAllImpl";
55}
56
57def AArch64O0PreLegalizerCombiner: GICombiner<
58  "AArch64O0PreLegalizerCombinerImpl", [optnone_combines]> {
59  let CombineAllMethodName = "tryCombineAllImpl";
60}
61
62// Matchdata for combines which replace a G_SHUFFLE_VECTOR with a
63// target-specific opcode.
64def shuffle_matchdata : GIDefMatchData<"ShuffleVectorPseudo">;
65
66def rev : GICombineRule<
67  (defs root:$root, shuffle_matchdata:$matchinfo),
68  (match (wip_match_opcode G_SHUFFLE_VECTOR):$root,
69         [{ return matchREV(*${root}, MRI, ${matchinfo}); }]),
70  (apply [{ applyShuffleVectorPseudo(*${root}, ${matchinfo}); }])
71>;
72
73def zip : GICombineRule<
74  (defs root:$root, shuffle_matchdata:$matchinfo),
75  (match (wip_match_opcode G_SHUFFLE_VECTOR):$root,
76         [{ return matchZip(*${root}, MRI, ${matchinfo}); }]),
77  (apply [{ applyShuffleVectorPseudo(*${root}, ${matchinfo}); }])
78>;
79
80def uzp : GICombineRule<
81  (defs root:$root, shuffle_matchdata:$matchinfo),
82  (match (wip_match_opcode G_SHUFFLE_VECTOR):$root,
83         [{ return matchUZP(*${root}, MRI, ${matchinfo}); }]),
84  (apply [{ applyShuffleVectorPseudo(*${root}, ${matchinfo}); }])
85>;
86
87def dup: GICombineRule <
88  (defs root:$root, shuffle_matchdata:$matchinfo),
89  (match (wip_match_opcode G_SHUFFLE_VECTOR):$root,
90         [{ return matchDup(*${root}, MRI, ${matchinfo}); }]),
91  (apply [{ applyShuffleVectorPseudo(*${root}, ${matchinfo}); }])
92>;
93
94def trn : GICombineRule<
95  (defs root:$root, shuffle_matchdata:$matchinfo),
96  (match (wip_match_opcode G_SHUFFLE_VECTOR):$root,
97         [{ return matchTRN(*${root}, MRI, ${matchinfo}); }]),
98  (apply [{ applyShuffleVectorPseudo(*${root}, ${matchinfo}); }])
99>;
100
101def ext: GICombineRule <
102  (defs root:$root, shuffle_matchdata:$matchinfo),
103  (match (wip_match_opcode G_SHUFFLE_VECTOR):$root,
104         [{ return matchEXT(*${root}, MRI, ${matchinfo}); }]),
105  (apply [{ applyEXT(*${root}, ${matchinfo}); }])
106>;
107
108def shuf_to_ins_matchdata : GIDefMatchData<"std::tuple<Register, int, Register, int>">;
109def shuf_to_ins: GICombineRule <
110  (defs root:$root, shuf_to_ins_matchdata:$matchinfo),
111  (match (wip_match_opcode G_SHUFFLE_VECTOR):$root,
112         [{ return matchINS(*${root}, MRI, ${matchinfo}); }]),
113  (apply [{ applyINS(*${root}, MRI, B, ${matchinfo}); }])
114>;
115
116def vashr_vlshr_imm_matchdata : GIDefMatchData<"int64_t">;
117def vashr_vlshr_imm : GICombineRule<
118  (defs root:$root, vashr_vlshr_imm_matchdata:$matchinfo),
119  (match (wip_match_opcode G_ASHR, G_LSHR):$root,
120          [{ return matchVAshrLshrImm(*${root}, MRI, ${matchinfo}); }]),
121  (apply [{ applyVAshrLshrImm(*${root}, MRI, ${matchinfo}); }])
122>;
123
124def form_duplane_matchdata :
125  GIDefMatchData<"std::pair<unsigned, int>">;
126def form_duplane : GICombineRule <
127  (defs root:$root, form_duplane_matchdata:$matchinfo),
128  (match (wip_match_opcode G_SHUFFLE_VECTOR):$root,
129          [{ return matchDupLane(*${root}, MRI, ${matchinfo}); }]),
130  (apply [{ applyDupLane(*${root}, MRI, B, ${matchinfo}); }])
131>;
132
133def shuffle_vector_lowering : GICombineGroup<[dup, rev, ext, zip, uzp, trn,
134                                              form_duplane,
135                                              shuf_to_ins]>;
136
137// Turn G_UNMERGE_VALUES -> G_EXTRACT_VECTOR_ELT's
138def vector_unmerge_lowering : GICombineRule <
139  (defs root:$root),
140  (match (wip_match_opcode G_UNMERGE_VALUES):$root,
141          [{ return matchScalarizeVectorUnmerge(*${root}, MRI); }]),
142  (apply [{ applyScalarizeVectorUnmerge(*${root}, MRI, B); }])
143>;
144
145def adjust_icmp_imm_matchdata :
146  GIDefMatchData<"std::pair<uint64_t, CmpInst::Predicate>">;
147def adjust_icmp_imm : GICombineRule <
148  (defs root:$root, adjust_icmp_imm_matchdata:$matchinfo),
149  (match (wip_match_opcode G_ICMP):$root,
150          [{ return matchAdjustICmpImmAndPred(*${root}, MRI, ${matchinfo}); }]),
151  (apply [{ applyAdjustICmpImmAndPred(*${root}, ${matchinfo}, B, Observer); }])
152>;
153
154def swap_icmp_operands : GICombineRule <
155  (defs root:$root),
156  (match (wip_match_opcode G_ICMP):$root,
157          [{ return trySwapICmpOperands(*${root}, MRI); }]),
158  (apply [{ applySwapICmpOperands(*${root}, Observer); }])
159>;
160
161def icmp_lowering : GICombineGroup<[adjust_icmp_imm, swap_icmp_operands]>;
162
163def extractvecelt_pairwise_add_matchdata : GIDefMatchData<"std::tuple<unsigned, LLT, Register>">;
164def extractvecelt_pairwise_add : GICombineRule<
165  (defs root:$root, extractvecelt_pairwise_add_matchdata:$matchinfo),
166  (match (wip_match_opcode G_EXTRACT_VECTOR_ELT):$root,
167          [{ return matchExtractVecEltPairwiseAdd(*${root}, MRI, ${matchinfo}); }]),
168  (apply [{ applyExtractVecEltPairwiseAdd(*${root}, MRI, B, ${matchinfo}); }])
169>;
170
171def mul_const_matchdata : GIDefMatchData<"std::function<void(MachineIRBuilder&, Register)>">;
172def mul_const : GICombineRule<
173  (defs root:$root, mul_const_matchdata:$matchinfo),
174  (match (wip_match_opcode G_MUL):$root,
175          [{ return matchAArch64MulConstCombine(*${root}, MRI, ${matchinfo}); }]),
176  (apply [{ applyAArch64MulConstCombine(*${root}, MRI, B, ${matchinfo}); }])
177>;
178
179def lower_mull : GICombineRule<
180  (defs root:$root),
181  (match (wip_match_opcode G_MUL):$root,
182          [{ return matchExtMulToMULL(*${root}, MRI); }]),
183  (apply [{ applyExtMulToMULL(*${root}, MRI, B, Observer); }])
184>;
185
186def build_vector_to_dup : GICombineRule<
187  (defs root:$root),
188  (match (wip_match_opcode G_BUILD_VECTOR):$root,
189          [{ return matchBuildVectorToDup(*${root}, MRI); }]),
190  (apply [{ applyBuildVectorToDup(*${root}, MRI, B); }])
191>;
192
193def build_vector_lowering : GICombineGroup<[build_vector_to_dup]>;
194
195def lower_vector_fcmp : GICombineRule<
196  (defs root:$root),
197  (match (wip_match_opcode G_FCMP):$root,
198    [{ return matchLowerVectorFCMP(*${root}, MRI, B); }]),
199  (apply [{ applyLowerVectorFCMP(*${root}, MRI, B); }])>;
200
201def form_truncstore_matchdata : GIDefMatchData<"Register">;
202def form_truncstore : GICombineRule<
203  (defs root:$root, form_truncstore_matchdata:$matchinfo),
204  (match (wip_match_opcode G_STORE):$root,
205          [{ return matchFormTruncstore(*${root}, MRI, ${matchinfo}); }]),
206  (apply [{ applyFormTruncstore(*${root}, MRI, B, Observer, ${matchinfo}); }])
207>;
208
209def fold_merge_to_zext : GICombineRule<
210  (defs root:$d),
211  (match (wip_match_opcode G_MERGE_VALUES):$d,
212          [{ return matchFoldMergeToZext(*${d}, MRI); }]),
213  (apply [{ applyFoldMergeToZext(*${d}, MRI, B, Observer); }])
214>;
215
216def mutate_anyext_to_zext : GICombineRule<
217  (defs root:$d),
218  (match (wip_match_opcode G_ANYEXT):$d,
219          [{ return matchMutateAnyExtToZExt(*${d}, MRI); }]),
220  (apply [{ applyMutateAnyExtToZExt(*${d}, MRI, B, Observer); }])
221>;
222
223def split_store_zero_128 : GICombineRule<
224  (defs root:$d),
225  (match (wip_match_opcode G_STORE):$d,
226          [{ return matchSplitStoreZero128(*${d}, MRI); }]),
227  (apply [{ applySplitStoreZero128(*${d}, MRI, B, Observer); }])
228>;
229
230def vector_sext_inreg_to_shift : GICombineRule<
231  (defs root:$d),
232  (match (wip_match_opcode G_SEXT_INREG):$d,
233          [{ return matchVectorSextInReg(*${d}, MRI); }]),
234  (apply [{ applyVectorSextInReg(*${d}, MRI, B, Observer); }])
235>;
236
237def unmerge_ext_to_unmerge_matchdata : GIDefMatchData<"Register">;
238def unmerge_ext_to_unmerge : GICombineRule<
239  (defs root:$d, unmerge_ext_to_unmerge_matchdata:$matchinfo),
240  (match (wip_match_opcode G_UNMERGE_VALUES):$d,
241          [{ return matchUnmergeExtToUnmerge(*${d}, MRI, ${matchinfo}); }]),
242  (apply [{ applyUnmergeExtToUnmerge(*${d}, MRI, B, Observer, ${matchinfo}); }])
243>;
244
245def regtriple_matchdata : GIDefMatchData<"std::tuple<Register, Register, Register>">;
246def or_to_bsp: GICombineRule <
247  (defs root:$root, regtriple_matchdata:$matchinfo),
248  (match (wip_match_opcode G_OR):$root,
249         [{ return matchOrToBSP(*${root}, MRI, ${matchinfo}); }]),
250  (apply [{ applyOrToBSP(*${root}, MRI, B, ${matchinfo}); }])
251>;
252
253// Post-legalization combines which should happen at all optimization levels.
254// (E.g. ones that facilitate matching for the selector) For example, matching
255// pseudos.
256def AArch64PostLegalizerLowering
257    : GICombiner<"AArch64PostLegalizerLoweringImpl",
258                       [shuffle_vector_lowering, vashr_vlshr_imm,
259                        icmp_lowering, build_vector_lowering,
260                        lower_vector_fcmp, form_truncstore,
261                        vector_sext_inreg_to_shift,
262                        unmerge_ext_to_unmerge, lower_mull,
263                        vector_unmerge_lowering]> {
264}
265
266// Post-legalization combines which are primarily optimizations.
267def AArch64PostLegalizerCombiner
268    : GICombiner<"AArch64PostLegalizerCombinerImpl",
269                       [copy_prop, combines_for_extload,
270                        combine_indexed_load_store,
271                        sext_trunc_sextload, mutate_anyext_to_zext,
272                        hoist_logic_op_with_same_opcode_hands,
273                        redundant_and, xor_of_and_with_same_reg,
274                        extractvecelt_pairwise_add, redundant_or,
275                        mul_const, redundant_sext_inreg,
276                        form_bitfield_extract, rotate_out_of_range,
277                        icmp_to_true_false_known_bits, merge_unmerge,
278                        select_combines, fold_merge_to_zext,
279                        constant_fold_binops, identity_combines,
280                        ptr_add_immed_chain, overlapping_and,
281                        split_store_zero_128, undef_combines,
282                        select_to_minmax, or_to_bsp]> {
283}
284