@@ -50,6 +50,12 @@ def GIAddrRegImm :
50
50
GIComplexOperandMatcher<s32, "selectAddrRegImm">,
51
51
GIComplexPatternEquiv<AddrRegImm>;
52
52
53
+ // Convert from i32 immediate to i64 target immediate to make SelectionDAG type
54
+ // checking happy so we can use ADDIW which expects an XLen immediate.
55
+ def as_i64imm : SDNodeXForm<imm, [{
56
+ return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i64);
57
+ }]>;
58
+
53
59
def gi_as_i64imm : GICustomOperandRenderer<"renderImm">,
54
60
GISDNodeXFormEquiv<as_i64imm>;
55
61
@@ -157,3 +163,305 @@ let Predicates = [IsRV64] in {
157
163
def : LdPat<load, LD, PtrVT>;
158
164
def : StPat<store, SD, GPR, PtrVT>;
159
165
}
166
+
167
+ //===----------------------------------------------------------------------===//
168
+ // RV64 i32 patterns not used by SelectionDAG
169
+ //===----------------------------------------------------------------------===//
170
+
171
+ def simm12i32 : ImmLeaf<i32, [{return isInt<12>(Imm);}]>;
172
+
173
+ def zext_is_sext : PatFrag<(ops node:$src), (zext node:$src), [{
174
+ KnownBits Known = CurDAG->computeKnownBits(N->getOperand(0), 0);
175
+ return Known.isNonNegative();
176
+ }]>;
177
+
178
+ let Predicates = [IsRV64] in {
179
+ def : LdPat<sextloadi8, LB, i32>;
180
+ def : LdPat<extloadi8, LBU, i32>; // Prefer unsigned due to no c.lb in Zcb.
181
+ def : LdPat<sextloadi16, LH, i32>;
182
+ def : LdPat<extloadi16, LH, i32>;
183
+ def : LdPat<zextloadi8, LBU, i32>;
184
+ def : LdPat<zextloadi16, LHU, i32>;
185
+
186
+ def : StPat<truncstorei8, SB, GPR, i32>;
187
+ def : StPat<truncstorei16, SH, GPR, i32>;
188
+
189
+ def : Pat<(anyext GPR:$src), (COPY GPR:$src)>;
190
+ def : Pat<(sext GPR:$src), (ADDIW GPR:$src, 0)>;
191
+ def : Pat<(trunc GPR:$src), (COPY GPR:$src)>;
192
+
193
+ def : PatGprGpr<add, ADDW, i32, i32>;
194
+ def : PatGprGpr<sub, SUBW, i32, i32>;
195
+ def : PatGprGpr<and, AND, i32, i32>;
196
+ def : PatGprGpr<or, OR, i32, i32>;
197
+ def : PatGprGpr<xor, XOR, i32, i32>;
198
+ def : PatGprGpr<shiftopw<shl>, SLLW, i32, i64>;
199
+ def : PatGprGpr<shiftopw<srl>, SRLW, i32, i64>;
200
+ def : PatGprGpr<shiftopw<sra>, SRAW, i32, i64>;
201
+
202
+ def : Pat<(i32 (add GPR:$rs1, simm12i32:$imm)),
203
+ (ADDIW GPR:$rs1, (i64 (as_i64imm $imm)))>;
204
+ def : Pat<(i32 (and GPR:$rs1, simm12i32:$imm)),
205
+ (ANDI GPR:$rs1, (i64 (as_i64imm $imm)))>;
206
+ def : Pat<(i32 (or GPR:$rs1, simm12i32:$imm)),
207
+ (ORI GPR:$rs1, (i64 (as_i64imm $imm)))>;
208
+ def : Pat<(i32 (xor GPR:$rs1, simm12i32:$imm)),
209
+ (XORI GPR:$rs1, (i64 (as_i64imm $imm)))>;
210
+
211
+ def : PatGprImm<shl, SLLIW, uimm5, i32>;
212
+ def : PatGprImm<srl, SRLIW, uimm5, i32>;
213
+ def : PatGprImm<sra, SRAIW, uimm5, i32>;
214
+
215
+ def : Pat<(i32 (and GPR:$rs, TrailingOnesMask:$mask)),
216
+ (SRLI (i32 (SLLI $rs, (i64 (XLenSubTrailingOnes $mask)))),
217
+ (i64 (XLenSubTrailingOnes $mask)))>;
218
+
219
+ // Use sext if the sign bit of the input is 0.
220
+ def : Pat<(zext_is_sext GPR:$src), (ADDIW GPR:$src, 0)>;
221
+ }
222
+
223
+ let Predicates = [IsRV64, NotHasStdExtZba] in {
224
+ def : Pat<(zext GPR:$src), (SRLI (i64 (SLLI GPR:$src, 32)), 32)>;
225
+
226
+ // If we're shifting a 32-bit zero extended value left by 0-31 bits, use 2
227
+ // shifts instead of 3. This can occur when unsigned is used to index an array.
228
+ def : Pat<(shl (zext GPR:$rs), uimm5:$shamt),
229
+ (SRLI (i64 (SLLI GPR:$rs, 32)), (ImmSubFrom32 uimm5:$shamt))>;
230
+ }
231
+
232
+ //===----------------------------------------------------------------------===//
233
+ // M RV64 i32 legalization patterns.
234
+ //===----------------------------------------------------------------------===//
235
+
236
+ let Predicates = [HasStdExtZmmul, IsRV64] in {
237
+ def : PatGprGpr<mul, MULW, i32, i32>;
238
+ }
239
+
240
+ let Predicates = [HasStdExtM, IsRV64] in {
241
+ def : PatGprGpr<sdiv, DIVW, i32, i32>;
242
+ def : PatGprGpr<udiv, DIVUW, i32, i32>;
243
+ def : PatGprGpr<srem, REMW, i32, i32>;
244
+ def : PatGprGpr<urem, REMUW, i32, i32>;
245
+ }
246
+
247
+ //===----------------------------------------------------------------------===//
248
+ // Atomic RV64 i32 patterns not used by SelectionDAG
249
+ //===----------------------------------------------------------------------===//
250
+
251
+ class PatGprGprA<SDPatternOperator OpNode, RVInst Inst, ValueType vt>
252
+ : Pat<(vt (OpNode (XLenVT GPR:$rs1), (vt GPR:$rs2))), (Inst GPR:$rs1, GPR:$rs2)>;
253
+
254
+ multiclass AMOPat2<string AtomicOp, string BaseInst, ValueType vt = XLenVT,
255
+ list<Predicate> ExtraPreds = []> {
256
+ let Predicates = !listconcat([HasStdExtA, NotHasStdExtZtso], ExtraPreds) in {
257
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_monotonic"),
258
+ !cast<RVInst>(BaseInst), vt>;
259
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_acquire"),
260
+ !cast<RVInst>(BaseInst#"_AQ"), vt>;
261
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_release"),
262
+ !cast<RVInst>(BaseInst#"_RL"), vt>;
263
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_acq_rel"),
264
+ !cast<RVInst>(BaseInst#"_AQ_RL"), vt>;
265
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_seq_cst"),
266
+ !cast<RVInst>(BaseInst#"_AQ_RL"), vt>;
267
+ }
268
+ let Predicates = !listconcat([HasStdExtA, HasStdExtZtso], ExtraPreds) in {
269
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_monotonic"),
270
+ !cast<RVInst>(BaseInst), vt>;
271
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_acquire"),
272
+ !cast<RVInst>(BaseInst), vt>;
273
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_release"),
274
+ !cast<RVInst>(BaseInst), vt>;
275
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_acq_rel"),
276
+ !cast<RVInst>(BaseInst), vt>;
277
+ def : PatGprGprA<!cast<PatFrag>(AtomicOp#"_seq_cst"),
278
+ !cast<RVInst>(BaseInst), vt>;
279
+ }
280
+ }
281
+
282
+ defm : AMOPat2<"atomic_swap_i32", "AMOSWAP_W", i32>;
283
+ defm : AMOPat2<"atomic_load_add_i32", "AMOADD_W", i32>;
284
+ defm : AMOPat2<"atomic_load_and_i32", "AMOAND_W", i32>;
285
+ defm : AMOPat2<"atomic_load_or_i32", "AMOOR_W", i32>;
286
+ defm : AMOPat2<"atomic_load_xor_i32", "AMOXOR_W", i32>;
287
+ defm : AMOPat2<"atomic_load_max_i32", "AMOMAX_W", i32>;
288
+ defm : AMOPat2<"atomic_load_min_i32", "AMOMIN_W", i32>;
289
+ defm : AMOPat2<"atomic_load_umax_i32", "AMOMAXU_W", i32>;
290
+ defm : AMOPat2<"atomic_load_umin_i32", "AMOMINU_W", i32>;
291
+
292
+ let Predicates = [HasStdExtA, IsRV64] in
293
+ defm : PseudoCmpXchgPat<"atomic_cmp_swap_i32", PseudoCmpXchg32, i32>;
294
+
295
+ let Predicates = [HasAtomicLdSt] in {
296
+ def : LdPat<atomic_load_8, LB, i32>;
297
+ def : LdPat<atomic_load_16, LH, i32>;
298
+ def : LdPat<atomic_load_32, LW, i32>;
299
+
300
+ def : StPat<atomic_store_8, SB, GPR, i32>;
301
+ def : StPat<atomic_store_16, SH, GPR, i32>;
302
+ def : StPat<atomic_store_32, SW, GPR, i32>;
303
+ }
304
+
305
+
306
+ //===----------------------------------------------------------------------===//
307
+ // Zb* RV64 i32 patterns not used by SelectionDAG.
308
+ //===----------------------------------------------------------------------===//
309
+
310
+ def zexti16i32 : ComplexPattern<i32, 1, "selectZExtBits<16>">;
311
+ def zexti8i32 : ComplexPattern<i32, 1, "selectZExtBits<8>">;
312
+
313
+ def BCLRMaski32 : ImmLeaf<i32, [{
314
+ return !isInt<12>(Imm) && isPowerOf2_32(~Imm);
315
+ }]>;
316
+ def SingleBitSetMaski32 : ImmLeaf<i32, [{
317
+ return !isInt<12>(Imm) && isPowerOf2_32(Imm);
318
+ }]>;
319
+
320
+ let Predicates = [HasStdExtZbb, IsRV64] in {
321
+ def : PatGpr<ctlz, CLZW, i32>;
322
+ def : PatGpr<cttz, CTZW, i32>;
323
+ def : PatGpr<ctpop, CPOPW, i32>;
324
+
325
+ def : Pat<(i32 (sext_inreg GPR:$rs1, i8)), (SEXT_B GPR:$rs1)>;
326
+ def : Pat<(i32 (sext_inreg GPR:$rs1, i16)), (SEXT_H GPR:$rs1)>;
327
+
328
+ def : Pat<(i32 (and GPR:$rs, 0xFFFF)), (ZEXT_H_RV64 GPR:$rs)>;
329
+ } // Predicates = [HasStdExtZbb, IsRV64]
330
+
331
+ let Predicates = [HasStdExtZbkb, NoStdExtZbb, IsRV64] in {
332
+ def : Pat<(i32 (and GPR:$rs, 0xFFFF)), (PACKW GPR:$rs, (XLenVT X0))>;
333
+ }
334
+
335
+ let Predicates = [HasStdExtZbbOrZbkb, IsRV64] in {
336
+ def : Pat<(i32 (and GPR:$rs1, (not GPR:$rs2))), (ANDN GPR:$rs1, GPR:$rs2)>;
337
+ def : Pat<(i32 (or GPR:$rs1, (not GPR:$rs2))), (ORN GPR:$rs1, GPR:$rs2)>;
338
+ def : Pat<(i32 (xor GPR:$rs1, (not GPR:$rs2))), (XNOR GPR:$rs1, GPR:$rs2)>;
339
+
340
+ def : PatGprGpr<shiftopw<rotl>, ROLW, i32, i64>;
341
+ def : PatGprGpr<shiftopw<rotr>, RORW, i32, i64>;
342
+ def : PatGprImm<rotr, RORIW, uimm5, i32>;
343
+
344
+ def : Pat<(i32 (rotl GPR:$rs1, uimm5:$rs2)),
345
+ (RORIW GPR:$rs1, (ImmSubFrom32 uimm5:$rs2))>;
346
+ } // Predicates = [HasStdExtZbbOrZbkb, IsRV64]
347
+
348
+ let Predicates = [HasStdExtZbkb, IsRV64] in {
349
+ def : Pat<(or (and (shl GPR:$rs2, (i64 8)), 0xFFFF),
350
+ (zexti8i32 (i32 GPR:$rs1))),
351
+ (PACKH GPR:$rs1, GPR:$rs2)>;
352
+ def : Pat<(or (shl (zexti8i32 (i32 GPR:$rs2)), (i64 8)),
353
+ (zexti8i32 (i32 GPR:$rs1))),
354
+ (PACKH GPR:$rs1, GPR:$rs2)>;
355
+ def : Pat<(and (anyext (or (shl GPR:$rs2, (XLenVT 8)),
356
+ (zexti8i32 (i32 GPR:$rs1)))), 0xFFFF),
357
+ (PACKH GPR:$rs1, GPR:$rs2)>;
358
+
359
+ def : Pat<(i32 (or (shl GPR:$rs2, (i64 16)), (zexti16i32 (i32 GPR:$rs1)))),
360
+ (PACKW GPR:$rs1, GPR:$rs2)>;
361
+ } // Predicates = [HasStdExtZbkb, IsRV64]
362
+
363
+ let Predicates = [HasStdExtZba, IsRV64] in {
364
+ def : Pat<(shl (i64 (zext i32:$rs1)), uimm5:$shamt),
365
+ (SLLI_UW GPR:$rs1, uimm5:$shamt)>;
366
+
367
+ def : Pat<(i64 (add_like_non_imm12 (zext GPR:$rs1), GPR:$rs2)),
368
+ (ADD_UW GPR:$rs1, GPR:$rs2)>;
369
+ def : Pat<(zext GPR:$src), (ADD_UW GPR:$src, (XLenVT X0))>;
370
+
371
+ foreach i = {1,2,3} in {
372
+ defvar shxadd = !cast<Instruction>("SH"#i#"ADD");
373
+ def : Pat<(i32 (add_like_non_imm12 (shl GPR:$rs1, (i64 i)), GPR:$rs2)),
374
+ (shxadd GPR:$rs1, GPR:$rs2)>;
375
+ def : Pat<(i32 (riscv_shl_add GPR:$rs1, (i32 i), GPR:$rs2)),
376
+ (shxadd GPR:$rs1, GPR:$rs2)>;
377
+ }
378
+ }
379
+
380
+ let Predicates = [HasStdExtZbs, IsRV64] in {
381
+ def : Pat<(i32 (and (not (shiftop<shl> 1, (i64 GPR:$rs2))), GPR:$rs1)),
382
+ (BCLR GPR:$rs1, GPR:$rs2)>;
383
+ def : Pat<(i32 (and (rotl -2, (i64 GPR:$rs2)), GPR:$rs1)),
384
+ (BCLR GPR:$rs1, GPR:$rs2)>;
385
+ def : Pat<(i32 (or (shiftop<shl> 1, (i64 GPR:$rs2)), GPR:$rs1)),
386
+ (BSET GPR:$rs1, GPR:$rs2)>;
387
+ def : Pat<(i32 (xor (shiftop<shl> 1, (i64 GPR:$rs2)), GPR:$rs1)),
388
+ (BINV GPR:$rs1, GPR:$rs2)>;
389
+ def : Pat<(i32 (and (shiftop<srl> GPR:$rs1, (i64 GPR:$rs2)), 1)),
390
+ (BEXT GPR:$rs1, GPR:$rs2)>;
391
+ def : Pat<(i64 (and (anyext (i32 (shiftop<srl> GPR:$rs1, (i64 GPR:$rs2)))), 1)),
392
+ (BEXT GPR:$rs1, GPR:$rs2)>;
393
+
394
+ def : Pat<(i32 (shiftop<shl> 1, (i64 GPR:$rs2))),
395
+ (BSET (XLenVT X0), GPR:$rs2)>;
396
+ def : Pat<(i32 (not (shiftop<shl> -1, (i64 GPR:$rs2)))),
397
+ (ADDI (i32 (BSET (XLenVT X0), GPR:$rs2)), -1)>;
398
+
399
+ def : Pat<(i32 (and (srl GPR:$rs1, uimm5:$shamt), (i32 1))),
400
+ (BEXTI GPR:$rs1, uimm5:$shamt)>;
401
+
402
+ def : Pat<(i32 (and GPR:$rs1, BCLRMaski32:$mask)),
403
+ (BCLRI GPR:$rs1, (i64 (BCLRXForm $mask)))>;
404
+ def : Pat<(i32 (or GPR:$rs1, SingleBitSetMaski32:$mask)),
405
+ (BSETI GPR:$rs1, (i64 (SingleBitSetMaskToIndex $mask)))>;
406
+ def : Pat<(i32 (xor GPR:$rs1, SingleBitSetMaski32:$mask)),
407
+ (BINVI GPR:$rs1, (i64 (SingleBitSetMaskToIndex $mask)))>;
408
+ } // Predicates = [HasStdExtZbs, IsRV64]
409
+
410
+ //===----------------------------------------------------------------------===//
411
+ // XTHead RV64 i32 patterns not used by SelectionDAG.
412
+ //===----------------------------------------------------------------------===//
413
+
414
+ def sexti16i32 : ComplexPattern<i32, 1, "selectSExtBits<16>">;
415
+
416
+ let Predicates = [HasVendorXTHeadMemIdx, IsRV64] in {
417
+ defm : StoreUpdatePat<post_truncsti8, TH_SBIA, i32>;
418
+ defm : StoreUpdatePat<pre_truncsti8, TH_SBIB, i32>;
419
+ defm : StoreUpdatePat<post_truncsti16, TH_SHIA, i32>;
420
+ defm : StoreUpdatePat<pre_truncsti16, TH_SHIB, i32>;
421
+
422
+ defm : StIdxPat<truncstorei8, TH_SRB, GPR, i32>;
423
+ defm : StIdxPat<truncstorei16, TH_SRH, GPR, i32>;
424
+
425
+ defm : StZextIdxPat<truncstorei8, TH_SURB, GPR, i32>;
426
+ defm : StZextIdxPat<truncstorei16, TH_SURH, GPR, i32>;
427
+ defm : StZextIdxPat<store, TH_SURW, GPR, i32>;
428
+ }
429
+
430
+ let Predicates = [HasVendorXTHeadCondMov, IsRV64] in {
431
+ def : Pat<(select (XLenVT GPR:$cond), (i32 GPR:$a), (i32 GPR:$b)),
432
+ (TH_MVEQZ GPR:$a, GPR:$b, GPR:$cond)>;
433
+ def : Pat<(select (XLenVT GPR:$cond), (i32 GPR:$a), (i32 0)),
434
+ (TH_MVEQZ GPR:$a, (XLenVT X0), GPR:$cond)>;
435
+ def : Pat<(select (XLenVT GPR:$cond), (i32 0), (i32 GPR:$b)),
436
+ (TH_MVNEZ GPR:$b, (XLenVT X0), GPR:$cond)>;
437
+
438
+ def : Pat<(select (riscv_seteq (XLenVT GPR:$cond)), (i32 GPR:$a), (i32 GPR:$b)),
439
+ (TH_MVNEZ GPR:$a, GPR:$b, GPR:$cond)>;
440
+ def : Pat<(select (riscv_setne (XLenVT GPR:$cond)), (i32 GPR:$a), (i32 GPR:$b)),
441
+ (TH_MVEQZ GPR:$a, GPR:$b, GPR:$cond)>;
442
+ def : Pat<(select (riscv_seteq (XLenVT GPR:$cond)), (i32 GPR:$a), (i32 0)),
443
+ (TH_MVNEZ GPR:$a, (XLenVT X0), GPR:$cond)>;
444
+ def : Pat<(select (riscv_setne (XLenVT GPR:$cond)), (i32 GPR:$a), (i32 0)),
445
+ (TH_MVEQZ GPR:$a, (XLenVT X0), GPR:$cond)>;
446
+ def : Pat<(select (riscv_seteq (XLenVT GPR:$cond)), (i32 0), (i32 GPR:$b)),
447
+ (TH_MVEQZ GPR:$b, (XLenVT X0), GPR:$cond)>;
448
+ def : Pat<(select (riscv_setne (XLenVT GPR:$cond)), (i32 0), (i32 GPR:$b)),
449
+ (TH_MVNEZ GPR:$b, (XLenVT X0), GPR:$cond)>;
450
+ } // Predicates = [HasVendorXTHeadCondMov]
451
+
452
+ let Predicates = [HasVendorXTHeadMac, IsRV64] in {
453
+ // mulaw, mulsw are available only in RV64.
454
+ def : Pat<(i32 (add GPR:$rd, (mul GPR:$rs1, GPR:$rs2))),
455
+ (TH_MULAW GPR:$rd, GPR:$rs1, GPR:$rs2)>;
456
+ def : Pat<(i32 (sub GPR:$rd, (mul GPR:$rs1, GPR:$rs2))),
457
+ (TH_MULSW GPR:$rd, GPR:$rs1, GPR:$rs2)>;
458
+ // mulah, mulsh produce a sign-extended result.
459
+ def : Pat<(i32 (add GPR:$rd,
460
+ (mul (sexti16i32 (i32 GPR:$rs1)),
461
+ (sexti16i32 (i32 GPR:$rs2))))),
462
+ (TH_MULAH GPR:$rd, GPR:$rs1, GPR:$rs2)>;
463
+ def : Pat<(i32 (sub GPR:$rd,
464
+ (mul (sexti16i32 (i32 GPR:$rs1)),
465
+ (sexti16i32 (i32 GPR:$rs2))))),
466
+ (TH_MULSH GPR:$rd, GPR:$rs1, GPR:$rs2)>;
467
+ }
0 commit comments