1 //===-- AMDGPUInstructions.td - Common instruction defs ---*- tablegen -*-===//
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
7 //===----------------------------------------------------------------------===//
9 // This file contains instruction defs that are common to all hw codegen
12 //===----------------------------------------------------------------------===//
14 class AddressSpacesImpl {
23 def AddrSpaces : AddressSpacesImpl;
26 class AMDGPUInst <dag outs, dag ins, string asm = "",
27 list<dag> pattern = []> : Instruction {
28 field bit isRegisterLoad = 0;
29 field bit isRegisterStore = 0;
31 let Namespace = "AMDGPU";
32 let OutOperandList = outs;
33 let InOperandList = ins;
35 let Pattern = pattern;
36 let Itinerary = NullALU;
38 // SoftFail is a field the disassembler can use to provide a way for
39 // instructions to not match without killing the whole decode process. It is
40 // mainly used for ARM, but Tablegen expects this field to exist or it fails
41 // to build the decode table.
42 field bits<64> SoftFail = 0;
44 let DecoderNamespace = Namespace;
46 let TSFlags{63} = isRegisterLoad;
47 let TSFlags{62} = isRegisterStore;
50 class AMDGPUShaderInst <dag outs, dag ins, string asm = "",
51 list<dag> pattern = []> : AMDGPUInst<outs, ins, asm, pattern> {
53 field bits<32> Inst = 0xffffffff;
56 //===---------------------------------------------------------------------===//
58 //===---------------------------------------------------------------------===//
60 class ILFormat<dag outs, dag ins, string asmstr, list<dag> pattern>
63 let Namespace = "AMDGPU";
64 dag OutOperandList = outs;
65 dag InOperandList = ins;
66 let Pattern = pattern;
67 let AsmString = !strconcat(asmstr, "\n");
69 let Itinerary = NullALU;
71 bit hasZeroOpFlag = 0;
74 let hasSideEffects = 0;
75 let isCodeGenOnly = 1;
78 def TruePredicate : Predicate<"">;
80 // FIXME: Tablegen should specially supports this
81 def FalsePredicate : Predicate<"false">;
83 // Add a predicate to the list if does not already exist to deduplicate it.
84 class PredConcat<list<Predicate> lst, Predicate pred> {
86 !foldl([pred], lst, acc, cur,
87 !listconcat(acc, !if(!eq(!cast<string>(cur),!cast<string>(pred)),
91 class PredicateControl {
92 Predicate SubtargetPredicate = TruePredicate;
93 Predicate AssemblerPredicate = TruePredicate;
94 Predicate WaveSizePredicate = TruePredicate;
95 list<Predicate> OtherPredicates = [];
96 list<Predicate> Predicates = PredConcat<
97 PredConcat<PredConcat<OtherPredicates,
98 SubtargetPredicate>.ret,
99 AssemblerPredicate>.ret,
100 WaveSizePredicate>.ret;
103 class AMDGPUPat<dag pattern, dag result> : Pat<pattern, result>,
106 let RecomputePerFunction = 1 in {
107 def FP16Denormals : Predicate<"MF->getInfo<SIMachineFunctionInfo>()->getMode().allFP64FP16Denormals()">;
108 def FP32Denormals : Predicate<"MF->getInfo<SIMachineFunctionInfo>()->getMode().allFP32Denormals()">;
109 def FP64Denormals : Predicate<"MF->getInfo<SIMachineFunctionInfo>()->getMode().allFP64FP16Denormals()">;
110 def NoFP16Denormals : Predicate<"!MF->getInfo<SIMachineFunctionInfo>()->getMode().allFP64FP16Denormals()">;
111 def NoFP32Denormals : Predicate<"!MF->getInfo<SIMachineFunctionInfo>()->getMode().allFP32Denormals()">;
112 def NoFP64Denormals : Predicate<"!MF->getInfo<SIMachineFunctionInfo>()->getMode().allFP64FP16Denormals()">;
113 def UnsafeFPMath : Predicate<"TM.Options.UnsafeFPMath">;
116 def FMA : Predicate<"Subtarget->hasFMA()">;
118 def InstFlag : OperandWithDefaultOps <i32, (ops (i32 0))>;
120 def u16ImmTarget : AsmOperandClass {
122 let RenderMethod = "addImmOperands";
125 def s16ImmTarget : AsmOperandClass {
127 let RenderMethod = "addImmOperands";
130 let OperandType = "OPERAND_IMMEDIATE" in {
132 def u32imm : Operand<i32> {
133 let PrintMethod = "printU32ImmOperand";
136 def u16imm : Operand<i16> {
137 let PrintMethod = "printU16ImmOperand";
138 let ParserMatchClass = u16ImmTarget;
141 def s16imm : Operand<i16> {
142 let PrintMethod = "printU16ImmOperand";
143 let ParserMatchClass = s16ImmTarget;
146 def u8imm : Operand<i8> {
147 let PrintMethod = "printU8ImmOperand";
150 } // End OperandType = "OPERAND_IMMEDIATE"
152 //===--------------------------------------------------------------------===//
154 //===--------------------------------------------------------------------===//
155 def brtarget : Operand<OtherVT>;
157 //===----------------------------------------------------------------------===//
159 //===----------------------------------------------------------------------===//
161 class HasOneUseUnaryOp<SDPatternOperator op> : PatFrag<
164 [{ return N->hasOneUse(); }]> {
166 let GISelPredicateCode = [{
167 return MRI.hasOneNonDBGUse(MI.getOperand(0).getReg());
171 class HasOneUseBinOp<SDPatternOperator op> : PatFrag<
172 (ops node:$src0, node:$src1),
174 [{ return N->hasOneUse(); }]> {
175 let GISelPredicateCode = [{
176 return MRI.hasOneNonDBGUse(MI.getOperand(0).getReg());
180 class HasOneUseTernaryOp<SDPatternOperator op> : PatFrag<
181 (ops node:$src0, node:$src1, node:$src2),
182 (op $src0, $src1, $src2),
183 [{ return N->hasOneUse(); }]> {
184 let GISelPredicateCode = [{
185 return MRI.hasOneNonDBGUse(MI.getOperand(0).getReg());
189 let Properties = [SDNPCommutative, SDNPAssociative] in {
190 def smax_oneuse : HasOneUseBinOp<smax>;
191 def smin_oneuse : HasOneUseBinOp<smin>;
192 def umax_oneuse : HasOneUseBinOp<umax>;
193 def umin_oneuse : HasOneUseBinOp<umin>;
195 def fminnum_oneuse : HasOneUseBinOp<fminnum>;
196 def fmaxnum_oneuse : HasOneUseBinOp<fmaxnum>;
198 def fminnum_ieee_oneuse : HasOneUseBinOp<fminnum_ieee>;
199 def fmaxnum_ieee_oneuse : HasOneUseBinOp<fmaxnum_ieee>;
202 def and_oneuse : HasOneUseBinOp<and>;
203 def or_oneuse : HasOneUseBinOp<or>;
204 def xor_oneuse : HasOneUseBinOp<xor>;
205 } // Properties = [SDNPCommutative, SDNPAssociative]
207 def not_oneuse : HasOneUseUnaryOp<not>;
209 def add_oneuse : HasOneUseBinOp<add>;
210 def sub_oneuse : HasOneUseBinOp<sub>;
212 def srl_oneuse : HasOneUseBinOp<srl>;
213 def shl_oneuse : HasOneUseBinOp<shl>;
215 def select_oneuse : HasOneUseTernaryOp<select>;
217 def AMDGPUmul_u24_oneuse : HasOneUseBinOp<AMDGPUmul_u24>;
218 def AMDGPUmul_i24_oneuse : HasOneUseBinOp<AMDGPUmul_i24>;
220 def srl_16 : PatFrag<
221 (ops node:$src0), (srl_oneuse node:$src0, (i32 16))
225 def hi_i16_elt : PatFrag<
226 (ops node:$src0), (i16 (trunc (i32 (srl_16 node:$src0))))
230 def hi_f16_elt : PatLeaf<
232 if (N->getOpcode() != ISD::BITCAST)
234 SDValue Tmp = N->getOperand(0);
236 if (Tmp.getOpcode() != ISD::SRL)
238 if (const auto *RHS = dyn_cast<ConstantSDNode>(Tmp.getOperand(1))
239 return RHS->getZExtValue() == 16;
243 //===----------------------------------------------------------------------===//
244 // PatLeafs for floating-point comparisons
245 //===----------------------------------------------------------------------===//
247 def COND_OEQ : PatFrags<(ops), [(OtherVT SETOEQ), (OtherVT SETEQ)]>;
248 def COND_ONE : PatFrags<(ops), [(OtherVT SETONE), (OtherVT SETNE)]>;
249 def COND_OGT : PatFrags<(ops), [(OtherVT SETOGT), (OtherVT SETGT)]>;
250 def COND_OGE : PatFrags<(ops), [(OtherVT SETOGE), (OtherVT SETGE)]>;
251 def COND_OLT : PatFrags<(ops), [(OtherVT SETOLT), (OtherVT SETLT)]>;
252 def COND_OLE : PatFrags<(ops), [(OtherVT SETOLE), (OtherVT SETLE)]>;
253 def COND_O : PatFrags<(ops), [(OtherVT SETO)]>;
254 def COND_UO : PatFrags<(ops), [(OtherVT SETUO)]>;
256 //===----------------------------------------------------------------------===//
257 // PatLeafs for unsigned / unordered comparisons
258 //===----------------------------------------------------------------------===//
260 def COND_UEQ : PatFrag<(ops), (OtherVT SETUEQ)>;
261 def COND_UNE : PatFrag<(ops), (OtherVT SETUNE)>;
262 def COND_UGT : PatFrag<(ops), (OtherVT SETUGT)>;
263 def COND_UGE : PatFrag<(ops), (OtherVT SETUGE)>;
264 def COND_ULT : PatFrag<(ops), (OtherVT SETULT)>;
265 def COND_ULE : PatFrag<(ops), (OtherVT SETULE)>;
267 // XXX - For some reason R600 version is preferring to use unordered
269 def COND_UNE_NE : PatFrags<(ops), [(OtherVT SETUNE), (OtherVT SETNE)]>;
271 //===----------------------------------------------------------------------===//
272 // PatLeafs for signed comparisons
273 //===----------------------------------------------------------------------===//
275 def COND_SGT : PatFrag<(ops), (OtherVT SETGT)>;
276 def COND_SGE : PatFrag<(ops), (OtherVT SETGE)>;
277 def COND_SLT : PatFrag<(ops), (OtherVT SETLT)>;
278 def COND_SLE : PatFrag<(ops), (OtherVT SETLE)>;
280 //===----------------------------------------------------------------------===//
281 // PatLeafs for integer equality
282 //===----------------------------------------------------------------------===//
284 def COND_EQ : PatFrags<(ops), [(OtherVT SETEQ), (OtherVT SETUEQ)]>;
285 def COND_NE : PatFrags<(ops), [(OtherVT SETNE), (OtherVT SETUNE)]>;
287 // FIXME: Should not need code predicate
288 //def COND_NULL : PatLeaf<(OtherVT null_frag)>;
289 def COND_NULL : PatLeaf <
291 [{(void)N; return false;}]
294 //===----------------------------------------------------------------------===//
295 // PatLeafs for Texture Constants
296 //===----------------------------------------------------------------------===//
298 def TEX_ARRAY : PatLeaf<
300 [{uint32_t TType = (uint32_t)N->getZExtValue();
301 return TType == 9 || TType == 10 || TType == 16;
305 def TEX_RECT : PatLeaf<
307 [{uint32_t TType = (uint32_t)N->getZExtValue();
312 def TEX_SHADOW : PatLeaf<
314 [{uint32_t TType = (uint32_t)N->getZExtValue();
315 return (TType >= 6 && TType <= 8) || TType == 13;
319 def TEX_SHADOW_ARRAY : PatLeaf<
321 [{uint32_t TType = (uint32_t)N->getZExtValue();
322 return TType == 11 || TType == 12 || TType == 17;
326 //===----------------------------------------------------------------------===//
327 // Load/Store Pattern Fragments
328 //===----------------------------------------------------------------------===//
330 def atomic_cmp_swap_glue : SDNode <"ISD::ATOMIC_CMP_SWAP", SDTAtomic3,
331 [SDNPHasChain, SDNPMayStore, SDNPMayLoad, SDNPMemOperand, SDNPInGlue]
334 class AddressSpaceList<list<int> AS> {
335 list<int> AddrSpaces = AS;
338 class Aligned<int Bytes> {
339 int MinAlignment = Bytes;
342 class StoreHi16<SDPatternOperator op> : PatFrag <
343 (ops node:$value, node:$ptr), (op (srl node:$value, (i32 16)), node:$ptr)> {
347 def LoadAddress_constant : AddressSpaceList<[ AddrSpaces.Constant ]>;
348 def LoadAddress_global : AddressSpaceList<[ AddrSpaces.Global, AddrSpaces.Constant ]>;
349 def StoreAddress_global : AddressSpaceList<[ AddrSpaces.Global ]>;
351 def LoadAddress_flat : AddressSpaceList<[ AddrSpaces.Flat,
353 AddrSpaces.Constant ]>;
354 def StoreAddress_flat : AddressSpaceList<[ AddrSpaces.Flat, AddrSpaces.Global ]>;
356 def LoadAddress_private : AddressSpaceList<[ AddrSpaces.Private ]>;
357 def StoreAddress_private : AddressSpaceList<[ AddrSpaces.Private ]>;
359 def LoadAddress_local : AddressSpaceList<[ AddrSpaces.Local ]>;
360 def StoreAddress_local : AddressSpaceList<[ AddrSpaces.Local ]>;
362 def LoadAddress_region : AddressSpaceList<[ AddrSpaces.Region ]>;
363 def StoreAddress_region : AddressSpaceList<[ AddrSpaces.Region ]>;
367 foreach as = [ "global", "flat", "constant", "local", "private", "region" ] in {
368 let AddressSpaces = !cast<AddressSpaceList>("LoadAddress_"#as).AddrSpaces in {
370 def load_#as : PatFrag<(ops node:$ptr), (unindexedload node:$ptr)> {
372 let IsNonExtLoad = 1;
375 def extloadi8_#as : PatFrag<(ops node:$ptr), (extload node:$ptr)> {
380 def extloadi16_#as : PatFrag<(ops node:$ptr), (extload node:$ptr)> {
385 def sextloadi8_#as : PatFrag<(ops node:$ptr), (sextload node:$ptr)> {
390 def sextloadi16_#as : PatFrag<(ops node:$ptr), (sextload node:$ptr)> {
395 def zextloadi8_#as : PatFrag<(ops node:$ptr), (zextload node:$ptr)> {
400 def zextloadi16_#as : PatFrag<(ops node:$ptr), (zextload node:$ptr)> {
405 def atomic_load_32_#as : PatFrag<(ops node:$ptr), (atomic_load_32 node:$ptr)> {
410 def atomic_load_64_#as : PatFrag<(ops node:$ptr), (atomic_load_64 node:$ptr)> {
414 } // End let AddressSpaces
418 foreach as = [ "global", "flat", "local", "private", "region" ] in {
419 let AddressSpaces = !cast<AddressSpaceList>("StoreAddress_"#as).AddrSpaces in {
420 def store_#as : PatFrag<(ops node:$val, node:$ptr),
421 (unindexedstore node:$val, node:$ptr)> {
423 let IsTruncStore = 0;
426 // truncstore fragments.
427 def truncstore_#as : PatFrag<(ops node:$val, node:$ptr),
428 (unindexedstore node:$val, node:$ptr)> {
430 let IsTruncStore = 1;
433 // TODO: We don't really need the truncstore here. We can use
434 // unindexedstore with MemoryVT directly, which will save an
435 // unnecessary check that the memory size is less than the value type
436 // in the generated matcher table.
437 def truncstorei8_#as : PatFrag<(ops node:$val, node:$ptr),
438 (truncstore node:$val, node:$ptr)> {
443 def truncstorei16_#as : PatFrag<(ops node:$val, node:$ptr),
444 (truncstore node:$val, node:$ptr)> {
449 def store_hi16_#as : StoreHi16 <truncstorei16>;
450 def truncstorei8_hi16_#as : StoreHi16<truncstorei8>;
451 def truncstorei16_hi16_#as : StoreHi16<truncstorei16>;
453 defm atomic_store_#as : binary_atomic_op<atomic_store>;
455 } // End let AddressSpaces
459 multiclass ret_noret_binary_atomic_op<SDNode atomic_op, bit IsInt = 1> {
460 foreach as = [ "global", "flat", "constant", "local", "private", "region" ] in {
461 let AddressSpaces = !cast<AddressSpaceList>("LoadAddress_"#as).AddrSpaces in {
462 defm "_"#as : binary_atomic_op<atomic_op, IsInt>;
464 let PredicateCode = [{return (SDValue(N, 0).use_empty());}] in {
465 defm "_"#as#"_noret" : binary_atomic_op<atomic_op, IsInt>;
468 let PredicateCode = [{return !(SDValue(N, 0).use_empty());}] in {
469 defm "_"#as#"_ret" : binary_atomic_op<atomic_op, IsInt>;
475 defm atomic_swap : ret_noret_binary_atomic_op<atomic_swap>;
476 defm atomic_load_add : ret_noret_binary_atomic_op<atomic_load_add>;
477 defm atomic_load_and : ret_noret_binary_atomic_op<atomic_load_and>;
478 defm atomic_load_max : ret_noret_binary_atomic_op<atomic_load_max>;
479 defm atomic_load_min : ret_noret_binary_atomic_op<atomic_load_min>;
480 defm atomic_load_or : ret_noret_binary_atomic_op<atomic_load_or>;
481 defm atomic_load_sub : ret_noret_binary_atomic_op<atomic_load_sub>;
482 defm atomic_load_umax : ret_noret_binary_atomic_op<atomic_load_umax>;
483 defm atomic_load_umin : ret_noret_binary_atomic_op<atomic_load_umin>;
484 defm atomic_load_xor : ret_noret_binary_atomic_op<atomic_load_xor>;
485 defm atomic_load_fadd : ret_noret_binary_atomic_op<atomic_load_fadd, 0>;
486 defm AMDGPUatomic_cmp_swap : ret_noret_binary_atomic_op<AMDGPUatomic_cmp_swap>;
489 def load_align8_local : PatFrag <(ops node:$ptr), (load_local node:$ptr)> {
491 let IsNonExtLoad = 1;
492 let MinAlignment = 8;
495 def load_align16_local : PatFrag <(ops node:$ptr), (load_local node:$ptr)> {
497 let IsNonExtLoad = 1;
498 let MinAlignment = 16;
501 def store_align8_local: PatFrag<(ops node:$val, node:$ptr),
502 (store_local node:$val, node:$ptr)>, Aligned<8> {
504 let IsTruncStore = 0;
507 def store_align16_local: PatFrag<(ops node:$val, node:$ptr),
508 (store_local node:$val, node:$ptr)>, Aligned<16> {
510 let IsTruncStore = 0;
513 let AddressSpaces = StoreAddress_local.AddrSpaces in {
514 defm atomic_cmp_swap_local : ternary_atomic_op<atomic_cmp_swap>;
515 defm atomic_cmp_swap_local_m0 : ternary_atomic_op<atomic_cmp_swap_glue>;
518 let AddressSpaces = StoreAddress_region.AddrSpaces in {
519 defm atomic_cmp_swap_region : ternary_atomic_op<atomic_cmp_swap>;
520 defm atomic_cmp_swap_region_m0 : ternary_atomic_op<atomic_cmp_swap_glue>;
523 //===----------------------------------------------------------------------===//
524 // Misc Pattern Fragments
525 //===----------------------------------------------------------------------===//
528 int TWO_PI = 0x40c90fdb;
530 int TWO_PI_INV = 0x3e22f983;
531 int FP_4294966784 = 0x4f7ffffe; // 4294966784 = 4294967296 - 512 = 2^32 - 2^9
532 int FP16_ONE = 0x3C00;
533 int FP16_NEG_ONE = 0xBC00;
534 int FP32_ONE = 0x3f800000;
535 int FP32_NEG_ONE = 0xbf800000;
536 int FP64_ONE = 0x3ff0000000000000;
537 int FP64_NEG_ONE = 0xbff0000000000000;
539 def CONST : Constants;
541 def FP_ZERO : PatLeaf <
543 [{return N->getValueAPF().isZero();}]
546 def FP_ONE : PatLeaf <
548 [{return N->isExactlyValue(1.0);}]
551 def FP_HALF : PatLeaf <
553 [{return N->isExactlyValue(0.5);}]
556 /* Generic helper patterns for intrinsics */
557 /* -------------------------------------- */
559 class POW_Common <AMDGPUInst log_ieee, AMDGPUInst exp_ieee, AMDGPUInst mul>
561 (fpow f32:$src0, f32:$src1),
562 (exp_ieee (mul f32:$src1, (log_ieee f32:$src0)))
565 /* Other helper patterns */
566 /* --------------------- */
568 /* Extract element pattern */
569 class Extract_Element <ValueType sub_type, ValueType vec_type, int sub_idx,
572 (sub_type (extractelt vec_type:$src, sub_idx)),
573 (EXTRACT_SUBREG $src, sub_reg)
576 /* Insert element pattern */
577 class Insert_Element <ValueType elem_type, ValueType vec_type,
578 int sub_idx, SubRegIndex sub_reg>
580 (insertelt vec_type:$vec, elem_type:$elem, sub_idx),
581 (INSERT_SUBREG $vec, $elem, sub_reg)
584 // XXX: Convert to new syntax and use COPY_TO_REG, once the DFAPacketizer
585 // can handle COPY instructions.
586 // bitconvert pattern
587 class BitConvert <ValueType dt, ValueType st, RegisterClass rc> : AMDGPUPat <
588 (dt (bitconvert (st rc:$src0))),
592 // XXX: Convert to new syntax and use COPY_TO_REG, once the DFAPacketizer
593 // can handle COPY instructions.
594 class DwordAddrPat<ValueType vt, RegisterClass rc> : AMDGPUPat <
595 (vt (AMDGPUdwordaddr (vt rc:$addr))),
601 multiclass BFIPatterns <Instruction BFI_INT,
602 Instruction LoadImm32,
603 RegisterClass RC64> {
604 // Definition from ISA doc:
605 // (y & x) | (z & ~x)
607 (or (and i32:$y, i32:$x), (and i32:$z, (not i32:$x))),
613 (or (and i64:$y, i64:$x), (and i64:$z, (not i64:$x))),
615 (BFI_INT (i32 (EXTRACT_SUBREG RC64:$x, sub0)),
616 (i32 (EXTRACT_SUBREG RC64:$y, sub0)),
617 (i32 (EXTRACT_SUBREG RC64:$z, sub0))), sub0,
618 (BFI_INT (i32 (EXTRACT_SUBREG RC64:$x, sub1)),
619 (i32 (EXTRACT_SUBREG RC64:$y, sub1)),
620 (i32 (EXTRACT_SUBREG RC64:$z, sub1))), sub1)
623 // SHA-256 Ch function
626 (xor i32:$z, (and i32:$x, (xor i32:$y, i32:$z))),
632 (xor i64:$z, (and i64:$x, (xor i64:$y, i64:$z))),
634 (BFI_INT (i32 (EXTRACT_SUBREG RC64:$x, sub0)),
635 (i32 (EXTRACT_SUBREG RC64:$y, sub0)),
636 (i32 (EXTRACT_SUBREG RC64:$z, sub0))), sub0,
637 (BFI_INT (i32 (EXTRACT_SUBREG RC64:$x, sub1)),
638 (i32 (EXTRACT_SUBREG RC64:$y, sub1)),
639 (i32 (EXTRACT_SUBREG RC64:$z, sub1))), sub1)
643 (fcopysign f32:$src0, f32:$src1),
644 (BFI_INT (LoadImm32 (i32 0x7fffffff)), $src0, $src1)
648 (f32 (fcopysign f32:$src0, f64:$src1)),
649 (BFI_INT (LoadImm32 (i32 0x7fffffff)), $src0,
650 (i32 (EXTRACT_SUBREG RC64:$src1, sub1)))
654 (f64 (fcopysign f64:$src0, f64:$src1)),
656 (i32 (EXTRACT_SUBREG $src0, sub0)), sub0,
657 (BFI_INT (LoadImm32 (i32 0x7fffffff)),
658 (i32 (EXTRACT_SUBREG RC64:$src0, sub1)),
659 (i32 (EXTRACT_SUBREG RC64:$src1, sub1))), sub1)
663 (f64 (fcopysign f64:$src0, f32:$src1)),
665 (i32 (EXTRACT_SUBREG $src0, sub0)), sub0,
666 (BFI_INT (LoadImm32 (i32 0x7fffffff)),
667 (i32 (EXTRACT_SUBREG RC64:$src0, sub1)),
672 // SHA-256 Ma patterns
674 // ((x & z) | (y & (x | z))) -> BFI_INT (XOR x, y), z, y
675 multiclass SHA256MaPattern <Instruction BFI_INT, Instruction XOR, RegisterClass RC64> {
677 (or (and i32:$x, i32:$z), (and i32:$y, (or i32:$x, i32:$z))),
678 (BFI_INT (XOR i32:$x, i32:$y), i32:$z, i32:$y)
682 (or (and i64:$x, i64:$z), (and i64:$y, (or i64:$x, i64:$z))),
684 (BFI_INT (XOR (i32 (EXTRACT_SUBREG RC64:$x, sub0)),
685 (i32 (EXTRACT_SUBREG RC64:$y, sub0))),
686 (i32 (EXTRACT_SUBREG RC64:$z, sub0)),
687 (i32 (EXTRACT_SUBREG RC64:$y, sub0))), sub0,
688 (BFI_INT (XOR (i32 (EXTRACT_SUBREG RC64:$x, sub1)),
689 (i32 (EXTRACT_SUBREG RC64:$y, sub1))),
690 (i32 (EXTRACT_SUBREG RC64:$z, sub1)),
691 (i32 (EXTRACT_SUBREG RC64:$y, sub1))), sub1)
695 // Bitfield extract patterns
697 def IMMZeroBasedBitfieldMask : ImmLeaf <i32, [{
698 return isMask_32(Imm);
701 def IMMPopCount : SDNodeXForm<imm, [{
702 return CurDAG->getTargetConstant(countPopulation(N->getZExtValue()), SDLoc(N),
706 multiclass BFEPattern <Instruction UBFE, Instruction SBFE, Instruction MOV> {
708 (i32 (and (i32 (srl i32:$src, i32:$rshift)), IMMZeroBasedBitfieldMask:$mask)),
709 (UBFE $src, $rshift, (MOV (i32 (IMMPopCount $mask))))
712 // x & ((1 << y) - 1)
714 (and i32:$src, (add_oneuse (shl_oneuse 1, i32:$width), -1)),
715 (UBFE $src, (MOV (i32 0)), $width)
720 (and i32:$src, (xor_oneuse (shl_oneuse -1, i32:$width), -1)),
721 (UBFE $src, (MOV (i32 0)), $width)
724 // x & (-1 >> (bitwidth - y))
726 (and i32:$src, (srl_oneuse -1, (sub 32, i32:$width))),
727 (UBFE $src, (MOV (i32 0)), $width)
730 // x << (bitwidth - y) >> (bitwidth - y)
732 (srl (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
733 (UBFE $src, (MOV (i32 0)), $width)
737 (sra (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
738 (SBFE $src, (MOV (i32 0)), $width)
743 class FSHRPattern <Instruction BIT_ALIGN> : AMDGPUPat <
744 (fshr i32:$src0, i32:$src1, i32:$src2),
745 (BIT_ALIGN $src0, $src1, $src2)
749 class ROTRPattern <Instruction BIT_ALIGN> : AMDGPUPat <
750 (rotr i32:$src0, i32:$src1),
751 (BIT_ALIGN $src0, $src0, $src1)
754 // Special conversion patterns
756 def cvt_rpi_i32_f32 : PatFrag <
758 (fp_to_sint (ffloor (fadd $src, FP_HALF))),
759 [{ (void) N; return TM.Options.NoNaNsFPMath; }]
762 def cvt_flr_i32_f32 : PatFrag <
764 (fp_to_sint (ffloor $src)),
765 [{ (void)N; return TM.Options.NoNaNsFPMath; }]
768 let AddedComplexity = 2 in {
769 class IMad24Pat<Instruction Inst, bit HasClamp = 0> : AMDGPUPat <
770 (add (AMDGPUmul_i24 i32:$src0, i32:$src1), i32:$src2),
771 !if(HasClamp, (Inst $src0, $src1, $src2, (i1 0)),
772 (Inst $src0, $src1, $src2))
775 class UMad24Pat<Instruction Inst, bit HasClamp = 0> : AMDGPUPat <
776 (add (AMDGPUmul_u24 i32:$src0, i32:$src1), i32:$src2),
777 !if(HasClamp, (Inst $src0, $src1, $src2, (i1 0)),
778 (Inst $src0, $src1, $src2))
780 } // AddedComplexity.
782 class RcpPat<Instruction RcpInst, ValueType vt> : AMDGPUPat <
783 (fdiv FP_ONE, vt:$src),
787 class RsqPat<Instruction RsqInst, ValueType vt> : AMDGPUPat <
788 (AMDGPUrcp (fsqrt vt:$src)),
792 // Instructions which select to the same v_min_f*
793 def fminnum_like : PatFrags<(ops node:$src0, node:$src1),
794 [(fminnum_ieee node:$src0, node:$src1),
795 (fminnum node:$src0, node:$src1)]
798 // Instructions which select to the same v_max_f*
799 def fmaxnum_like : PatFrags<(ops node:$src0, node:$src1),
800 [(fmaxnum_ieee node:$src0, node:$src1),
801 (fmaxnum node:$src0, node:$src1)]
804 def fminnum_like_oneuse : PatFrags<(ops node:$src0, node:$src1),
805 [(fminnum_ieee_oneuse node:$src0, node:$src1),
806 (fminnum_oneuse node:$src0, node:$src1)]
809 def fmaxnum_like_oneuse : PatFrags<(ops node:$src0, node:$src1),
810 [(fmaxnum_ieee_oneuse node:$src0, node:$src1),
811 (fmaxnum_oneuse node:$src0, node:$src1)]
814 def any_fmad : PatFrags<(ops node:$src0, node:$src1, node:$src2),
815 [(fmad node:$src0, node:$src1, node:$src2),
816 (AMDGPUfmad_ftz node:$src0, node:$src1, node:$src2)]
819 // FIXME: fsqrt should not select directly
820 def any_amdgcn_sqrt : PatFrags<(ops node:$src0),
821 [(fsqrt node:$src0), (int_amdgcn_sqrt node:$src0)]