1 //===- X86LegalizerInfo.cpp --------------------------------------*- C++ -*-==//
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 /// \file
9 /// This file implements the targeting of the Machinelegalizer class for X86.
10 /// \todo This should be generated by TableGen.
11 //===----------------------------------------------------------------------===//
12
13 #include "X86LegalizerInfo.h"
14 #include "X86Subtarget.h"
15 #include "X86TargetMachine.h"
16 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
17 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
18 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
19 #include "llvm/CodeGen/MachineConstantPool.h"
20 #include "llvm/CodeGen/TargetOpcodes.h"
21 #include "llvm/CodeGen/ValueTypes.h"
22 #include "llvm/IR/DerivedTypes.h"
23 #include "llvm/IR/Type.h"
24
25 using namespace llvm;
26 using namespace TargetOpcode;
27 using namespace LegalizeActions;
28 using namespace LegalityPredicates;
29
X86LegalizerInfo(const X86Subtarget & STI,const X86TargetMachine & TM)30 X86LegalizerInfo::X86LegalizerInfo(const X86Subtarget &STI,
31 const X86TargetMachine &TM)
32 : Subtarget(STI) {
33
34 bool Is64Bit = Subtarget.is64Bit();
35 bool HasCMOV = Subtarget.canUseCMOV();
36 bool HasSSE1 = Subtarget.hasSSE1();
37 bool HasSSE2 = Subtarget.hasSSE2();
38 bool HasSSE41 = Subtarget.hasSSE41();
39 bool HasAVX = Subtarget.hasAVX();
40 bool HasAVX2 = Subtarget.hasAVX2();
41 bool HasAVX512 = Subtarget.hasAVX512();
42 bool HasVLX = Subtarget.hasVLX();
43 bool HasDQI = Subtarget.hasAVX512() && Subtarget.hasDQI();
44 bool HasBWI = Subtarget.hasAVX512() && Subtarget.hasBWI();
45 bool UseX87 = !Subtarget.useSoftFloat() && Subtarget.hasX87();
46
47 const LLT p0 = LLT::pointer(0, TM.getPointerSizeInBits(0));
48 const LLT s1 = LLT::scalar(1);
49 const LLT s8 = LLT::scalar(8);
50 const LLT s16 = LLT::scalar(16);
51 const LLT s32 = LLT::scalar(32);
52 const LLT s64 = LLT::scalar(64);
53 const LLT s80 = LLT::scalar(80);
54 const LLT s128 = LLT::scalar(128);
55 const LLT sMaxScalar = Subtarget.is64Bit() ? s64 : s32;
56 const LLT v2s32 = LLT::fixed_vector(2, 32);
57 const LLT v4s8 = LLT::fixed_vector(4, 8);
58
59
60 const LLT v16s8 = LLT::fixed_vector(16, 8);
61 const LLT v8s16 = LLT::fixed_vector(8, 16);
62 const LLT v4s32 = LLT::fixed_vector(4, 32);
63 const LLT v2s64 = LLT::fixed_vector(2, 64);
64 const LLT v2p0 = LLT::fixed_vector(2, p0);
65
66 const LLT v32s8 = LLT::fixed_vector(32, 8);
67 const LLT v16s16 = LLT::fixed_vector(16, 16);
68 const LLT v8s32 = LLT::fixed_vector(8, 32);
69 const LLT v4s64 = LLT::fixed_vector(4, 64);
70 const LLT v4p0 = LLT::fixed_vector(4, p0);
71
72 const LLT v64s8 = LLT::fixed_vector(64, 8);
73 const LLT v32s16 = LLT::fixed_vector(32, 16);
74 const LLT v16s32 = LLT::fixed_vector(16, 32);
75 const LLT v8s64 = LLT::fixed_vector(8, 64);
76
77 const LLT s8MaxVector = HasAVX512 ? v64s8 : HasAVX ? v32s8 : v16s8;
78 const LLT s16MaxVector = HasAVX512 ? v32s16 : HasAVX ? v16s16 : v8s16;
79 const LLT s32MaxVector = HasAVX512 ? v16s32 : HasAVX ? v8s32 : v4s32;
80 const LLT s64MaxVector = HasAVX512 ? v8s64 : HasAVX ? v4s64 : v2s64;
81
82 // todo: AVX512 bool vector predicate types
83
84 // implicit/constants
85 getActionDefinitionsBuilder(G_IMPLICIT_DEF)
86 .legalIf([=](const LegalityQuery &Query) -> bool {
87 // 32/64-bits needs support for s64/s128 to handle cases:
88 // s64 = EXTEND (G_IMPLICIT_DEF s32) -> s64 = G_IMPLICIT_DEF
89 // s128 = EXTEND (G_IMPLICIT_DEF s32/s64) -> s128 = G_IMPLICIT_DEF
90 return typeInSet(0, {p0, s1, s8, s16, s32, s64})(Query) ||
91 (Is64Bit && typeInSet(0, {s128})(Query));
92 });
93
94 getActionDefinitionsBuilder(G_CONSTANT)
95 .legalIf([=](const LegalityQuery &Query) -> bool {
96 return typeInSet(0, {p0, s8, s16, s32})(Query) ||
97 (Is64Bit && typeInSet(0, {s64})(Query));
98 })
99 .widenScalarToNextPow2(0, /*Min=*/8)
100 .clampScalar(0, s8, sMaxScalar);
101
102 // merge/unmerge
103 for (unsigned Op : {G_MERGE_VALUES, G_UNMERGE_VALUES}) {
104 unsigned BigTyIdx = Op == G_MERGE_VALUES ? 0 : 1;
105 unsigned LitTyIdx = Op == G_MERGE_VALUES ? 1 : 0;
106 getActionDefinitionsBuilder(Op)
107 .widenScalarToNextPow2(LitTyIdx, /*Min=*/8)
108 .widenScalarToNextPow2(BigTyIdx, /*Min=*/16)
109 .minScalar(LitTyIdx, s8)
110 .minScalar(BigTyIdx, s32)
111 .legalIf([=](const LegalityQuery &Q) {
112 switch (Q.Types[BigTyIdx].getSizeInBits()) {
113 case 16:
114 case 32:
115 case 64:
116 case 128:
117 case 256:
118 case 512:
119 break;
120 default:
121 return false;
122 }
123 switch (Q.Types[LitTyIdx].getSizeInBits()) {
124 case 8:
125 case 16:
126 case 32:
127 case 64:
128 case 128:
129 case 256:
130 return true;
131 default:
132 return false;
133 }
134 });
135 }
136
137 // integer addition/subtraction
138 getActionDefinitionsBuilder({G_ADD, G_SUB})
139 .legalIf([=](const LegalityQuery &Query) -> bool {
140 if (typeInSet(0, {s8, s16, s32})(Query))
141 return true;
142 if (Is64Bit && typeInSet(0, {s64})(Query))
143 return true;
144 if (HasSSE2 && typeInSet(0, {v16s8, v8s16, v4s32, v2s64})(Query))
145 return true;
146 if (HasAVX2 && typeInSet(0, {v32s8, v16s16, v8s32, v4s64})(Query))
147 return true;
148 if (HasAVX512 && typeInSet(0, {v16s32, v8s64})(Query))
149 return true;
150 if (HasBWI && typeInSet(0, {v64s8, v32s16})(Query))
151 return true;
152 return false;
153 })
154 .clampMinNumElements(0, s8, 16)
155 .clampMinNumElements(0, s16, 8)
156 .clampMinNumElements(0, s32, 4)
157 .clampMinNumElements(0, s64, 2)
158 .clampMaxNumElements(0, s8, HasBWI ? 64 : (HasAVX2 ? 32 : 16))
159 .clampMaxNumElements(0, s16, HasBWI ? 32 : (HasAVX2 ? 16 : 8))
160 .clampMaxNumElements(0, s32, HasAVX512 ? 16 : (HasAVX2 ? 8 : 4))
161 .clampMaxNumElements(0, s64, HasAVX512 ? 8 : (HasAVX2 ? 4 : 2))
162 .widenScalarToNextPow2(0, /*Min=*/32)
163 .clampScalar(0, s8, sMaxScalar)
164 .scalarize(0);
165
166 getActionDefinitionsBuilder({G_UADDE, G_UADDO, G_USUBE, G_USUBO})
167 .legalIf([=](const LegalityQuery &Query) -> bool {
168 return typePairInSet(0, 1, {{s8, s1}, {s16, s1}, {s32, s1}})(Query) ||
169 (Is64Bit && typePairInSet(0, 1, {{s64, s1}})(Query));
170 })
171 .widenScalarToNextPow2(0, /*Min=*/32)
172 .clampScalar(0, s8, sMaxScalar)
173 .clampScalar(1, s1, s1)
174 .scalarize(0);
175
176 // integer multiply
177 getActionDefinitionsBuilder(G_MUL)
178 .legalIf([=](const LegalityQuery &Query) -> bool {
179 if (typeInSet(0, {s8, s16, s32})(Query))
180 return true;
181 if (Is64Bit && typeInSet(0, {s64})(Query))
182 return true;
183 if (HasSSE2 && typeInSet(0, {v8s16})(Query))
184 return true;
185 if (HasSSE41 && typeInSet(0, {v4s32})(Query))
186 return true;
187 if (HasAVX2 && typeInSet(0, {v16s16, v8s32})(Query))
188 return true;
189 if (HasAVX512 && typeInSet(0, {v16s32})(Query))
190 return true;
191 if (HasDQI && typeInSet(0, {v8s64})(Query))
192 return true;
193 if (HasDQI && HasVLX && typeInSet(0, {v2s64, v4s64})(Query))
194 return true;
195 if (HasBWI && typeInSet(0, {v32s16})(Query))
196 return true;
197 return false;
198 })
199 .clampMinNumElements(0, s16, 8)
200 .clampMinNumElements(0, s32, 4)
201 .clampMinNumElements(0, s64, HasVLX ? 2 : 8)
202 .clampMaxNumElements(0, s16, HasBWI ? 32 : (HasAVX2 ? 16 : 8))
203 .clampMaxNumElements(0, s32, HasAVX512 ? 16 : (HasAVX2 ? 8 : 4))
204 .clampMaxNumElements(0, s64, 8)
205 .widenScalarToNextPow2(0, /*Min=*/32)
206 .clampScalar(0, s8, sMaxScalar)
207 .scalarize(0);
208
209 getActionDefinitionsBuilder({G_SMULH, G_UMULH})
210 .legalIf([=](const LegalityQuery &Query) -> bool {
211 return typeInSet(0, {s8, s16, s32})(Query) ||
212 (Is64Bit && typeInSet(0, {s64})(Query));
213 })
214 .widenScalarToNextPow2(0, /*Min=*/32)
215 .clampScalar(0, s8, sMaxScalar)
216 .scalarize(0);
217
218 // integer divisions
219 getActionDefinitionsBuilder({G_SDIV, G_SREM, G_UDIV, G_UREM})
220 .legalIf([=](const LegalityQuery &Query) -> bool {
221 return typeInSet(0, {s8, s16, s32})(Query) ||
222 (Is64Bit && typeInSet(0, {s64})(Query));
223 })
224 .libcallFor({s64})
225 .clampScalar(0, s8, sMaxScalar);
226
227 // integer shifts
228 getActionDefinitionsBuilder({G_SHL, G_LSHR, G_ASHR})
229 .legalIf([=](const LegalityQuery &Query) -> bool {
230 return typePairInSet(0, 1, {{s8, s8}, {s16, s8}, {s32, s8}})(Query) ||
231 (Is64Bit && typePairInSet(0, 1, {{s64, s8}})(Query));
232 })
233 .clampScalar(0, s8, sMaxScalar)
234 .clampScalar(1, s8, s8);
235
236 // integer logic
237 getActionDefinitionsBuilder({G_AND, G_OR, G_XOR})
238 .legalIf([=](const LegalityQuery &Query) -> bool {
239 if (typeInSet(0, {s8, s16, s32})(Query))
240 return true;
241 if (Is64Bit && typeInSet(0, {s64})(Query))
242 return true;
243 if (HasSSE2 && typeInSet(0, {v16s8, v8s16, v4s32, v2s64})(Query))
244 return true;
245 if (HasAVX && typeInSet(0, {v32s8, v16s16, v8s32, v4s64})(Query))
246 return true;
247 if (HasAVX512 && typeInSet(0, {v64s8, v32s16, v16s32, v8s64})(Query))
248 return true;
249 return false;
250 })
251 .clampMinNumElements(0, s8, 16)
252 .clampMinNumElements(0, s16, 8)
253 .clampMinNumElements(0, s32, 4)
254 .clampMinNumElements(0, s64, 2)
255 .clampMaxNumElements(0, s8, HasAVX512 ? 64 : (HasAVX ? 32 : 16))
256 .clampMaxNumElements(0, s16, HasAVX512 ? 32 : (HasAVX ? 16 : 8))
257 .clampMaxNumElements(0, s32, HasAVX512 ? 16 : (HasAVX ? 8 : 4))
258 .clampMaxNumElements(0, s64, HasAVX512 ? 8 : (HasAVX ? 4 : 2))
259 .widenScalarToNextPow2(0, /*Min=*/32)
260 .clampScalar(0, s8, sMaxScalar)
261 .scalarize(0);
262
263 // integer comparison
264 const std::initializer_list<LLT> IntTypes32 = {s8, s16, s32, p0};
265 const std::initializer_list<LLT> IntTypes64 = {s8, s16, s32, s64, p0};
266
267 getActionDefinitionsBuilder(G_ICMP)
268 .legalForCartesianProduct({s8}, Is64Bit ? IntTypes64 : IntTypes32)
269 .clampScalar(0, s8, s8)
270 .clampScalar(1, s8, sMaxScalar);
271
272 // bswap
273 getActionDefinitionsBuilder(G_BSWAP)
274 .legalIf([=](const LegalityQuery &Query) {
275 return Query.Types[0] == s32 ||
276 (Subtarget.is64Bit() && Query.Types[0] == s64);
277 })
278 .widenScalarToNextPow2(0, /*Min=*/32)
279 .clampScalar(0, s32, sMaxScalar);
280
281 // popcount
282 getActionDefinitionsBuilder(G_CTPOP)
283 .legalIf([=](const LegalityQuery &Query) -> bool {
284 return Subtarget.hasPOPCNT() &&
285 (typePairInSet(0, 1, {{s16, s16}, {s32, s32}})(Query) ||
286 (Is64Bit && typePairInSet(0, 1, {{s64, s64}})(Query)));
287 })
288 .widenScalarToNextPow2(1, /*Min=*/16)
289 .clampScalar(1, s16, sMaxScalar)
290 .scalarSameSizeAs(0, 1);
291
292 // count leading zeros (LZCNT)
293 getActionDefinitionsBuilder(G_CTLZ)
294 .legalIf([=](const LegalityQuery &Query) -> bool {
295 return Subtarget.hasLZCNT() &&
296 (typePairInSet(0, 1, {{s16, s16}, {s32, s32}})(Query) ||
297 (Is64Bit && typePairInSet(0, 1, {{s64, s64}})(Query)));
298 })
299 .widenScalarToNextPow2(1, /*Min=*/16)
300 .clampScalar(1, s16, sMaxScalar)
301 .scalarSameSizeAs(0, 1);
302
303 // count trailing zeros
304 getActionDefinitionsBuilder({G_CTTZ_ZERO_UNDEF, G_CTTZ})
305 .legalIf([=](const LegalityQuery &Query) -> bool {
306 return (Query.Opcode == G_CTTZ_ZERO_UNDEF || Subtarget.hasBMI()) &&
307 (typePairInSet(0, 1, {{s16, s16}, {s32, s32}})(Query) ||
308 (Is64Bit && typePairInSet(0, 1, {{s64, s64}})(Query)));
309 })
310 .widenScalarToNextPow2(1, /*Min=*/16)
311 .clampScalar(1, s16, sMaxScalar)
312 .scalarSameSizeAs(0, 1);
313
314 // control flow
315 getActionDefinitionsBuilder(G_PHI)
316 .legalIf([=](const LegalityQuery &Query) -> bool {
317 return typeInSet(0, {s8, s16, s32, p0})(Query) ||
318 (Is64Bit && typeInSet(0, {s64})(Query)) ||
319 (HasSSE1 && typeInSet(0, {v16s8, v8s16, v4s32, v2s64})(Query)) ||
320 (HasAVX && typeInSet(0, {v32s8, v16s16, v8s32, v4s64})(Query)) ||
321 (HasAVX512 &&
322 typeInSet(0, {v64s8, v32s16, v16s32, v8s64})(Query));
323 })
324 .clampMinNumElements(0, s8, 16)
325 .clampMinNumElements(0, s16, 8)
326 .clampMinNumElements(0, s32, 4)
327 .clampMinNumElements(0, s64, 2)
328 .clampMaxNumElements(0, s8, HasAVX512 ? 64 : (HasAVX ? 32 : 16))
329 .clampMaxNumElements(0, s16, HasAVX512 ? 32 : (HasAVX ? 16 : 8))
330 .clampMaxNumElements(0, s32, HasAVX512 ? 16 : (HasAVX ? 8 : 4))
331 .clampMaxNumElements(0, s64, HasAVX512 ? 8 : (HasAVX ? 4 : 2))
332 .widenScalarToNextPow2(0, /*Min=*/32)
333 .clampScalar(0, s8, sMaxScalar)
334 .scalarize(0);
335
336 getActionDefinitionsBuilder(G_BRCOND).legalFor({s1});
337
338 // pointer handling
339 const std::initializer_list<LLT> PtrTypes32 = {s1, s8, s16, s32};
340 const std::initializer_list<LLT> PtrTypes64 = {s1, s8, s16, s32, s64};
341
342 getActionDefinitionsBuilder(G_PTRTOINT)
343 .legalForCartesianProduct(Is64Bit ? PtrTypes64 : PtrTypes32, {p0})
344 .maxScalar(0, sMaxScalar)
345 .widenScalarToNextPow2(0, /*Min*/ 8);
346
347 getActionDefinitionsBuilder(G_INTTOPTR).legalFor({{p0, sMaxScalar}});
348
349 getActionDefinitionsBuilder(G_CONSTANT_POOL).legalFor({p0});
350
351 getActionDefinitionsBuilder(G_PTR_ADD)
352 .legalIf([=](const LegalityQuery &Query) -> bool {
353 return typePairInSet(0, 1, {{p0, s32}})(Query) ||
354 (Is64Bit && typePairInSet(0, 1, {{p0, s64}})(Query));
355 })
356 .widenScalarToNextPow2(1, /*Min*/ 32)
357 .clampScalar(1, s32, sMaxScalar);
358
359 getActionDefinitionsBuilder({G_FRAME_INDEX, G_GLOBAL_VALUE}).legalFor({p0});
360
361 // load/store: add more corner cases
362 for (unsigned Op : {G_LOAD, G_STORE}) {
363 auto &Action = getActionDefinitionsBuilder(Op);
364 Action.legalForTypesWithMemDesc({{s8, p0, s1, 1},
365 {s8, p0, s8, 1},
366 {s16, p0, s8, 1},
367 {s16, p0, s16, 1},
368 {s32, p0, s8, 1},
369 {s32, p0, s16, 1},
370 {s32, p0, s32, 1},
371 {s80, p0, s80, 1},
372 {p0, p0, p0, 1},
373 {v4s8, p0, v4s8, 1}});
374 if (Is64Bit)
375 Action.legalForTypesWithMemDesc({{s64, p0, s8, 1},
376 {s64, p0, s16, 1},
377 {s64, p0, s32, 1},
378 {s64, p0, s64, 1},
379 {v2s32, p0, v2s32, 1}});
380 if (HasSSE1)
381 Action.legalForTypesWithMemDesc({{v4s32, p0, v4s32, 1}});
382 if (HasSSE2)
383 Action.legalForTypesWithMemDesc({{v16s8, p0, v16s8, 1},
384 {v8s16, p0, v8s16, 1},
385 {v2s64, p0, v2s64, 1},
386 {v2p0, p0, v2p0, 1}});
387 if (HasAVX)
388 Action.legalForTypesWithMemDesc({{v32s8, p0, v32s8, 1},
389 {v16s16, p0, v16s16, 1},
390 {v8s32, p0, v8s32, 1},
391 {v4s64, p0, v4s64, 1},
392 {v4p0, p0, v4p0, 1}});
393 if (HasAVX512)
394 Action.legalForTypesWithMemDesc({{v64s8, p0, v64s8, 1},
395 {v32s16, p0, v32s16, 1},
396 {v16s32, p0, v16s32, 1},
397 {v8s64, p0, v8s64, 1}});
398 Action.widenScalarToNextPow2(0, /*Min=*/8)
399 .clampScalar(0, s8, sMaxScalar)
400 .scalarize(0);
401 }
402
403 for (unsigned Op : {G_SEXTLOAD, G_ZEXTLOAD}) {
404 auto &Action = getActionDefinitionsBuilder(Op);
405 Action.legalForTypesWithMemDesc({{s16, p0, s8, 1},
406 {s32, p0, s8, 1},
407 {s32, p0, s16, 1}});
408 if (Is64Bit)
409 Action.legalForTypesWithMemDesc({{s64, p0, s8, 1},
410 {s64, p0, s16, 1},
411 {s64, p0, s32, 1}});
412 // TODO - SSE41/AVX2/AVX512F/AVX512BW vector extensions
413 }
414
415 // sext, zext, and anyext
416 getActionDefinitionsBuilder({G_SEXT, G_ZEXT, G_ANYEXT})
417 .legalIf([=](const LegalityQuery &Query) {
418 return typeInSet(0, {s8, s16, s32})(Query) ||
419 (Query.Opcode == G_ANYEXT && Query.Types[0] == s128) ||
420 (Is64Bit && Query.Types[0] == s64);
421 })
422 .widenScalarToNextPow2(0, /*Min=*/8)
423 .clampScalar(0, s8, sMaxScalar)
424 .widenScalarToNextPow2(1, /*Min=*/8)
425 .clampScalar(1, s8, sMaxScalar)
426 .scalarize(0);
427
428 getActionDefinitionsBuilder(G_SEXT_INREG).lower();
429
430 // fp constants
431 getActionDefinitionsBuilder(G_FCONSTANT)
432 .legalIf([=](const LegalityQuery &Query) -> bool {
433 return (typeInSet(0, {s32, s64})(Query)) ||
434 (UseX87 && typeInSet(0, {s80})(Query));
435 });
436
437 // fp arithmetic
438 getActionDefinitionsBuilder({G_FADD, G_FSUB, G_FMUL, G_FDIV})
439 .legalIf([=](const LegalityQuery &Query) {
440 return (typeInSet(0, {s32, s64})(Query)) ||
441 (HasSSE1 && typeInSet(0, {v4s32})(Query)) ||
442 (HasSSE2 && typeInSet(0, {v2s64})(Query)) ||
443 (HasAVX && typeInSet(0, {v8s32, v4s64})(Query)) ||
444 (HasAVX512 && typeInSet(0, {v16s32, v8s64})(Query)) ||
445 (UseX87 && typeInSet(0, {s80})(Query));
446 });
447
448 // fp comparison
449 getActionDefinitionsBuilder(G_FCMP)
450 .legalIf([=](const LegalityQuery &Query) {
451 return (HasSSE1 && typePairInSet(0, 1, {{s8, s32}})(Query)) ||
452 (HasSSE2 && typePairInSet(0, 1, {{s8, s64}})(Query));
453 })
454 .clampScalar(0, s8, s8)
455 .clampScalar(1, s32, HasSSE2 ? s64 : s32)
456 .widenScalarToNextPow2(1);
457
458 // fp conversions
459 getActionDefinitionsBuilder(G_FPEXT).legalIf([=](const LegalityQuery &Query) {
460 return (HasSSE2 && typePairInSet(0, 1, {{s64, s32}})(Query)) ||
461 (HasAVX && typePairInSet(0, 1, {{v4s64, v4s32}})(Query)) ||
462 (HasAVX512 && typePairInSet(0, 1, {{v8s64, v8s32}})(Query));
463 });
464
465 getActionDefinitionsBuilder(G_FPTRUNC).legalIf(
466 [=](const LegalityQuery &Query) {
467 return (HasSSE2 && typePairInSet(0, 1, {{s32, s64}})(Query)) ||
468 (HasAVX && typePairInSet(0, 1, {{v4s32, v4s64}})(Query)) ||
469 (HasAVX512 && typePairInSet(0, 1, {{v8s32, v8s64}})(Query));
470 });
471
472 getActionDefinitionsBuilder(G_SITOFP)
473 .legalIf([=](const LegalityQuery &Query) {
474 return (HasSSE1 &&
475 (typePairInSet(0, 1, {{s32, s32}})(Query) ||
476 (Is64Bit && typePairInSet(0, 1, {{s32, s64}})(Query)))) ||
477 (HasSSE2 &&
478 (typePairInSet(0, 1, {{s64, s32}})(Query) ||
479 (Is64Bit && typePairInSet(0, 1, {{s64, s64}})(Query))));
480 })
481 .clampScalar(1, s32, sMaxScalar)
482 .widenScalarToNextPow2(1)
483 .clampScalar(0, s32, HasSSE2 ? s64 : s32)
484 .widenScalarToNextPow2(0);
485
486 getActionDefinitionsBuilder(G_FPTOSI)
487 .legalIf([=](const LegalityQuery &Query) {
488 return (HasSSE1 &&
489 (typePairInSet(0, 1, {{s32, s32}})(Query) ||
490 (Is64Bit && typePairInSet(0, 1, {{s64, s32}})(Query)))) ||
491 (HasSSE2 &&
492 (typePairInSet(0, 1, {{s32, s64}})(Query) ||
493 (Is64Bit && typePairInSet(0, 1, {{s64, s64}})(Query))));
494 })
495 .clampScalar(1, s32, HasSSE2 ? s64 : s32)
496 .widenScalarToNextPow2(0)
497 .clampScalar(0, s32, sMaxScalar)
498 .widenScalarToNextPow2(1);
499
500 // vector ops
501 getActionDefinitionsBuilder(G_BUILD_VECTOR)
502 .customIf([=](const LegalityQuery &Query) {
503 return (HasSSE1 && typeInSet(0, {v4s32})(Query)) ||
504 (HasSSE2 && typeInSet(0, {v2s64, v8s16, v16s8})(Query)) ||
505 (HasAVX && typeInSet(0, {v4s64, v8s32, v16s16, v32s8})(Query)) ||
506 (HasAVX512 && typeInSet(0, {v8s64, v16s32, v32s16, v64s8}));
507 })
508 .clampNumElements(0, v16s8, s8MaxVector)
509 .clampNumElements(0, v8s16, s16MaxVector)
510 .clampNumElements(0, v4s32, s32MaxVector)
511 .clampNumElements(0, v2s64, s64MaxVector)
512 .moreElementsToNextPow2(0);
513
514 getActionDefinitionsBuilder({G_EXTRACT, G_INSERT})
515 .legalIf([=](const LegalityQuery &Query) {
516 unsigned SubIdx = Query.Opcode == G_EXTRACT ? 0 : 1;
517 unsigned FullIdx = Query.Opcode == G_EXTRACT ? 1 : 0;
518 return (HasAVX && typePairInSet(SubIdx, FullIdx,
519 {{v16s8, v32s8},
520 {v8s16, v16s16},
521 {v4s32, v8s32},
522 {v2s64, v4s64}})(Query)) ||
523 (HasAVX512 && typePairInSet(SubIdx, FullIdx,
524 {{v16s8, v64s8},
525 {v32s8, v64s8},
526 {v8s16, v32s16},
527 {v16s16, v32s16},
528 {v4s32, v16s32},
529 {v8s32, v16s32},
530 {v2s64, v8s64},
531 {v4s64, v8s64}})(Query));
532 });
533
534 // todo: only permit dst types up to max legal vector register size?
535 getActionDefinitionsBuilder(G_CONCAT_VECTORS)
536 .legalIf([=](const LegalityQuery &Query) {
537 return (HasSSE1 && typePairInSet(1, 0,
538 {{v16s8, v32s8},
539 {v8s16, v16s16},
540 {v4s32, v8s32},
541 {v2s64, v4s64}})(Query)) ||
542 (HasAVX && typePairInSet(1, 0,
543 {{v16s8, v64s8},
544 {v32s8, v64s8},
545 {v8s16, v32s16},
546 {v16s16, v32s16},
547 {v4s32, v16s32},
548 {v8s32, v16s32},
549 {v2s64, v8s64},
550 {v4s64, v8s64}})(Query));
551 });
552
553 // todo: vectors and address spaces
554 getActionDefinitionsBuilder(G_SELECT)
555 .legalFor({{s8, s32}, {s16, s32}, {s32, s32}, {s64, s32}, {p0, s32}})
556 .widenScalarToNextPow2(0, /*Min=*/8)
557 .clampScalar(0, HasCMOV ? s16 : s8, sMaxScalar)
558 .clampScalar(1, s32, s32);
559
560 // memory intrinsics
561 getActionDefinitionsBuilder({G_MEMCPY, G_MEMMOVE, G_MEMSET}).libcall();
562
563 getActionDefinitionsBuilder({G_DYN_STACKALLOC,
564 G_STACKSAVE,
565 G_STACKRESTORE}).lower();
566
567 // fp intrinsics
568 getActionDefinitionsBuilder(G_INTRINSIC_ROUNDEVEN)
569 .scalarize(0)
570 .minScalar(0, LLT::scalar(32))
571 .libcall();
572
573 getActionDefinitionsBuilder({G_FREEZE, G_CONSTANT_FOLD_BARRIER})
574 .legalFor({s8, s16, s32, s64, p0})
575 .widenScalarToNextPow2(0, /*Min=*/8)
576 .clampScalar(0, s8, sMaxScalar);
577
578 getLegacyLegalizerInfo().computeTables();
579 verify(*STI.getInstrInfo());
580 }
581
legalizeCustom(LegalizerHelper & Helper,MachineInstr & MI,LostDebugLocObserver & LocObserver) const582 bool X86LegalizerInfo::legalizeCustom(LegalizerHelper &Helper, MachineInstr &MI,
583 LostDebugLocObserver &LocObserver) const {
584 MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
585 MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
586 switch (MI.getOpcode()) {
587 default:
588 // No idea what to do.
589 return false;
590 case TargetOpcode::G_BUILD_VECTOR:
591 return legalizeBuildVector(MI, MRI, Helper);
592 }
593 llvm_unreachable("expected switch to return");
594 }
595
legalizeBuildVector(MachineInstr & MI,MachineRegisterInfo & MRI,LegalizerHelper & Helper) const596 bool X86LegalizerInfo::legalizeBuildVector(MachineInstr &MI,
597 MachineRegisterInfo &MRI,
598 LegalizerHelper &Helper) const {
599 MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
600 const auto &BuildVector = cast<GBuildVector>(MI);
601 Register Dst = BuildVector.getReg(0);
602 LLT DstTy = MRI.getType(Dst);
603 MachineFunction &MF = MIRBuilder.getMF();
604 LLVMContext &Ctx = MF.getFunction().getContext();
605 uint64_t DstTySize = DstTy.getScalarSizeInBits();
606
607 SmallVector<Constant *, 4> CstIdxs;
608 for (unsigned i = 0; i < BuildVector.getNumSources(); ++i) {
609 Register Source = BuildVector.getSourceReg(i);
610
611 auto ValueAndReg = getIConstantVRegValWithLookThrough(Source, MRI);
612 if (ValueAndReg) {
613 CstIdxs.emplace_back(ConstantInt::get(Ctx, ValueAndReg->Value));
614 continue;
615 }
616
617 auto FPValueAndReg = getFConstantVRegValWithLookThrough(Source, MRI);
618 if (FPValueAndReg) {
619 CstIdxs.emplace_back(ConstantFP::get(Ctx, FPValueAndReg->Value));
620 continue;
621 }
622
623 if (getOpcodeDef<GImplicitDef>(Source, MRI)) {
624 CstIdxs.emplace_back(UndefValue::get(Type::getIntNTy(Ctx, DstTySize)));
625 continue;
626 }
627 return false;
628 }
629
630 Constant *ConstVal = ConstantVector::get(CstIdxs);
631
632 const DataLayout &DL = MIRBuilder.getDataLayout();
633 unsigned AddrSpace = DL.getDefaultGlobalsAddressSpace();
634 Align Alignment(DL.getABITypeAlign(ConstVal->getType()));
635 auto Addr = MIRBuilder.buildConstantPool(
636 LLT::pointer(AddrSpace, DL.getPointerSizeInBits(AddrSpace)),
637 MF.getConstantPool()->getConstantPoolIndex(ConstVal, Alignment));
638 MachineMemOperand *MMO =
639 MF.getMachineMemOperand(MachinePointerInfo::getConstantPool(MF),
640 MachineMemOperand::MOLoad, DstTy, Alignment);
641
642 MIRBuilder.buildLoad(Dst, Addr, *MMO);
643 MI.eraseFromParent();
644 return true;
645 }
646
legalizeIntrinsic(LegalizerHelper & Helper,MachineInstr & MI) const647 bool X86LegalizerInfo::legalizeIntrinsic(LegalizerHelper &Helper,
648 MachineInstr &MI) const {
649 return true;
650 }
651