1 //===-- SIInstrInfo.td - SI Instruction Infos -------------*- 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 def isWave32 : Predicate<"Subtarget->getWavefrontSize() == 32">,
10 AssemblerPredicate <"FeatureWavefrontSize32">;
11 def isWave64 : Predicate<"Subtarget->getWavefrontSize() == 64">,
12 AssemblerPredicate <"FeatureWavefrontSize64">;
14 def DisableInst : Predicate <"false">, AssemblerPredicate<"FeatureDisable">;
16 class GCNPredicateControl : PredicateControl {
17 Predicate SIAssemblerPredicate = isGFX6GFX7;
18 Predicate VIAssemblerPredicate = isGFX8GFX9;
21 // Execpt for the NONE field, this must be kept in sync with the
22 // SIEncodingFamily enum in AMDGPUInstrInfo.cpp
23 def SIEncodingFamily {
35 //===----------------------------------------------------------------------===//
37 //===----------------------------------------------------------------------===//
39 def AMDGPUclamp : SDNode<"AMDGPUISD::CLAMP", SDTFPUnaryOp>;
41 def SIsbuffer_load : SDNode<"AMDGPUISD::SBUFFER_LOAD",
42 SDTypeProfile<1, 4, [SDTCisVT<1, v4i32>, SDTCisVT<2, i32>, SDTCisVT<3, i1>,
44 [SDNPMayLoad, SDNPMemOperand]
47 def SIds_ordered_count : SDNode<"AMDGPUISD::DS_ORDERED_COUNT",
48 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisVT<2, i16>]>,
49 [SDNPMayLoad, SDNPMayStore, SDNPMemOperand, SDNPHasChain, SDNPInGlue]
52 def SIatomic_inc : SDNode<"AMDGPUISD::ATOMIC_INC", SDTAtomic2,
53 [SDNPMayLoad, SDNPMayStore, SDNPMemOperand, SDNPHasChain]
56 def SIatomic_dec : SDNode<"AMDGPUISD::ATOMIC_DEC", SDTAtomic2,
57 [SDNPMayLoad, SDNPMayStore, SDNPMemOperand, SDNPHasChain]
60 def SDTAtomic2_f32 : SDTypeProfile<1, 2, [
61 SDTCisSameAs<0,2>, SDTCisFP<0>, SDTCisPtrTy<1>
64 def SIatomic_fmin : SDNode<"AMDGPUISD::ATOMIC_LOAD_FMIN", SDTAtomic2_f32,
65 [SDNPMayLoad, SDNPMayStore, SDNPMemOperand, SDNPHasChain]
68 def SIatomic_fmax : SDNode<"AMDGPUISD::ATOMIC_LOAD_FMAX", SDTAtomic2_f32,
69 [SDNPMayLoad, SDNPMayStore, SDNPMemOperand, SDNPHasChain]
72 // load_d16_{lo|hi} ptr, tied_input
73 def SIload_d16 : SDTypeProfile<1, 2, [
79 def SDTtbuffer_load : SDTypeProfile<1, 8,
81 SDTCisVT<1, v4i32>, // rsrc
82 SDTCisVT<2, i32>, // vindex(VGPR)
83 SDTCisVT<3, i32>, // voffset(VGPR)
84 SDTCisVT<4, i32>, // soffset(SGPR)
85 SDTCisVT<5, i32>, // offset(imm)
86 SDTCisVT<6, i32>, // format(imm)
87 SDTCisVT<7, i32>, // cachecontrol(imm)
88 SDTCisVT<8, i1> // idxen(imm)
91 def SItbuffer_load : SDNode<"AMDGPUISD::TBUFFER_LOAD_FORMAT", SDTtbuffer_load,
92 [SDNPMayLoad, SDNPMemOperand, SDNPHasChain]>;
93 def SItbuffer_load_d16 : SDNode<"AMDGPUISD::TBUFFER_LOAD_FORMAT_D16",
95 [SDNPMayLoad, SDNPMemOperand, SDNPHasChain]>;
97 def SDTtbuffer_store : SDTypeProfile<0, 9,
99 SDTCisVT<1, v4i32>, // rsrc
100 SDTCisVT<2, i32>, // vindex(VGPR)
101 SDTCisVT<3, i32>, // voffset(VGPR)
102 SDTCisVT<4, i32>, // soffset(SGPR)
103 SDTCisVT<5, i32>, // offset(imm)
104 SDTCisVT<6, i32>, // format(imm)
105 SDTCisVT<7, i32>, // cachecontrol(imm)
106 SDTCisVT<8, i1> // idxen(imm)
109 def SItbuffer_store : SDNode<"AMDGPUISD::TBUFFER_STORE_FORMAT", SDTtbuffer_store,
110 [SDNPMayStore, SDNPMemOperand, SDNPHasChain]>;
111 def SItbuffer_store_d16 : SDNode<"AMDGPUISD::TBUFFER_STORE_FORMAT_D16",
113 [SDNPMayStore, SDNPMemOperand, SDNPHasChain]>;
115 def SDTBufferLoad : SDTypeProfile<1, 7,
117 SDTCisVT<1, v4i32>, // rsrc
118 SDTCisVT<2, i32>, // vindex(VGPR)
119 SDTCisVT<3, i32>, // voffset(VGPR)
120 SDTCisVT<4, i32>, // soffset(SGPR)
121 SDTCisVT<5, i32>, // offset(imm)
122 SDTCisVT<6, i32>, // cachepolicy(imm)
123 SDTCisVT<7, i1>]>; // idxen(imm)
125 def SIbuffer_load : SDNode <"AMDGPUISD::BUFFER_LOAD", SDTBufferLoad,
126 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
127 def SIbuffer_load_ubyte : SDNode <"AMDGPUISD::BUFFER_LOAD_UBYTE", SDTBufferLoad,
128 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
129 def SIbuffer_load_ushort : SDNode <"AMDGPUISD::BUFFER_LOAD_USHORT", SDTBufferLoad,
130 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
131 def SIbuffer_load_byte : SDNode <"AMDGPUISD::BUFFER_LOAD_BYTE", SDTBufferLoad,
132 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
133 def SIbuffer_load_short: SDNode <"AMDGPUISD::BUFFER_LOAD_SHORT", SDTBufferLoad,
134 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
135 def SIbuffer_load_format : SDNode <"AMDGPUISD::BUFFER_LOAD_FORMAT", SDTBufferLoad,
136 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
137 def SIbuffer_load_format_d16 : SDNode <"AMDGPUISD::BUFFER_LOAD_FORMAT_D16",
139 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
141 def SDTBufferStore : SDTypeProfile<0, 8,
143 SDTCisVT<1, v4i32>, // rsrc
144 SDTCisVT<2, i32>, // vindex(VGPR)
145 SDTCisVT<3, i32>, // voffset(VGPR)
146 SDTCisVT<4, i32>, // soffset(SGPR)
147 SDTCisVT<5, i32>, // offset(imm)
148 SDTCisVT<6, i32>, // cachepolicy(imm)
149 SDTCisVT<7, i1>]>; // idxen(imm)
151 def SIbuffer_store : SDNode <"AMDGPUISD::BUFFER_STORE", SDTBufferStore,
152 [SDNPMayStore, SDNPMemOperand, SDNPHasChain]>;
153 def SIbuffer_store_byte: SDNode <"AMDGPUISD::BUFFER_STORE_BYTE",
155 [SDNPMayStore, SDNPMemOperand, SDNPHasChain]>;
156 def SIbuffer_store_short : SDNode <"AMDGPUISD::BUFFER_STORE_SHORT",
158 [SDNPMayStore, SDNPMemOperand, SDNPHasChain]>;
159 def SIbuffer_store_format : SDNode <"AMDGPUISD::BUFFER_STORE_FORMAT",
161 [SDNPMayStore, SDNPMemOperand, SDNPHasChain]>;
162 def SIbuffer_store_format_d16 : SDNode <"AMDGPUISD::BUFFER_STORE_FORMAT_D16",
164 [SDNPMayStore, SDNPMemOperand, SDNPHasChain]>;
166 class SDBufferAtomic<string opcode> : SDNode <opcode,
168 [SDTCisVT<2, v4i32>, // rsrc
169 SDTCisVT<3, i32>, // vindex(VGPR)
170 SDTCisVT<4, i32>, // voffset(VGPR)
171 SDTCisVT<5, i32>, // soffset(SGPR)
172 SDTCisVT<6, i32>, // offset(imm)
173 SDTCisVT<7, i32>, // cachepolicy(imm)
174 SDTCisVT<8, i1>]>, // idxen(imm)
175 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad, SDNPMayStore]
178 class SDBufferAtomicNoRtn<string opcode, ValueType ty> : SDNode <opcode,
180 [SDTCisVT<0, ty>, // vdata
181 SDTCisVT<1, v4i32>, // rsrc
182 SDTCisVT<2, i32>, // vindex(VGPR)
183 SDTCisVT<3, i32>, // voffset(VGPR)
184 SDTCisVT<4, i32>, // soffset(SGPR)
185 SDTCisVT<5, i32>, // offset(imm)
186 SDTCisVT<6, i32>, // cachepolicy(imm)
187 SDTCisVT<7, i1>]>, // idxen(imm)
188 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad, SDNPMayStore]
191 def SIbuffer_atomic_swap : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_SWAP">;
192 def SIbuffer_atomic_add : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_ADD">;
193 def SIbuffer_atomic_sub : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_SUB">;
194 def SIbuffer_atomic_smin : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_SMIN">;
195 def SIbuffer_atomic_umin : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_UMIN">;
196 def SIbuffer_atomic_smax : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_SMAX">;
197 def SIbuffer_atomic_umax : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_UMAX">;
198 def SIbuffer_atomic_and : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_AND">;
199 def SIbuffer_atomic_or : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_OR">;
200 def SIbuffer_atomic_xor : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_XOR">;
201 def SIbuffer_atomic_inc : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_INC">;
202 def SIbuffer_atomic_dec : SDBufferAtomic <"AMDGPUISD::BUFFER_ATOMIC_DEC">;
203 def SIbuffer_atomic_fadd : SDBufferAtomicNoRtn <"AMDGPUISD::BUFFER_ATOMIC_FADD", f32>;
204 def SIbuffer_atomic_pk_fadd : SDBufferAtomicNoRtn <"AMDGPUISD::BUFFER_ATOMIC_PK_FADD", v2f16>;
206 def SIbuffer_atomic_cmpswap : SDNode <"AMDGPUISD::BUFFER_ATOMIC_CMPSWAP",
208 [SDTCisVT<0, i32>, // dst
209 SDTCisVT<1, i32>, // src
210 SDTCisVT<2, i32>, // cmp
211 SDTCisVT<3, v4i32>, // rsrc
212 SDTCisVT<4, i32>, // vindex(VGPR)
213 SDTCisVT<5, i32>, // voffset(VGPR)
214 SDTCisVT<6, i32>, // soffset(SGPR)
215 SDTCisVT<7, i32>, // offset(imm)
216 SDTCisVT<8, i32>, // cachepolicy(imm)
217 SDTCisVT<9, i1>]>, // idxen(imm)
218 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad, SDNPMayStore]
221 class SDGlobalAtomicNoRtn<string opcode, ValueType ty> : SDNode <opcode,
223 [SDTCisPtrTy<0>, // vaddr
224 SDTCisVT<1, ty>]>, // vdata
225 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad, SDNPMayStore]
228 def SIglobal_atomic_fadd : SDGlobalAtomicNoRtn <"AMDGPUISD::ATOMIC_FADD", f32>;
229 def SIglobal_atomic_pk_fadd : SDGlobalAtomicNoRtn <"AMDGPUISD::ATOMIC_PK_FADD", v2f16>;
231 def SIpc_add_rel_offset : SDNode<"AMDGPUISD::PC_ADD_REL_OFFSET",
232 SDTypeProfile<1, 2, [SDTCisVT<0, iPTR>, SDTCisSameAs<0,1>, SDTCisSameAs<0,2>]>
235 def SIlds : SDNode<"AMDGPUISD::LDS",
236 SDTypeProfile<1, 1, [SDTCisVT<0, iPTR>, SDTCisSameAs<0,1>]>
239 def SIload_d16_lo : SDNode<"AMDGPUISD::LOAD_D16_LO",
241 [SDNPMayLoad, SDNPMemOperand, SDNPHasChain]
244 def SIload_d16_lo_u8 : SDNode<"AMDGPUISD::LOAD_D16_LO_U8",
246 [SDNPMayLoad, SDNPMemOperand, SDNPHasChain]
249 def SIload_d16_lo_i8 : SDNode<"AMDGPUISD::LOAD_D16_LO_I8",
251 [SDNPMayLoad, SDNPMemOperand, SDNPHasChain]
254 def SIload_d16_hi : SDNode<"AMDGPUISD::LOAD_D16_HI",
256 [SDNPMayLoad, SDNPMemOperand, SDNPHasChain]
259 def SIload_d16_hi_u8 : SDNode<"AMDGPUISD::LOAD_D16_HI_U8",
261 [SDNPMayLoad, SDNPMemOperand, SDNPHasChain]
264 def SIload_d16_hi_i8 : SDNode<"AMDGPUISD::LOAD_D16_HI_I8",
266 [SDNPMayLoad, SDNPMemOperand, SDNPHasChain]
269 def SIdenorm_mode : SDNode<"AMDGPUISD::DENORM_MODE",
270 SDTypeProfile<0 ,1, [SDTCisInt<0>]>,
271 [SDNPHasChain, SDNPSideEffect, SDNPOptInGlue, SDNPOutGlue]
274 //===----------------------------------------------------------------------===//
276 //===----------------------------------------------------------------------===//
278 // Returns 1 if the source arguments have modifiers, 0 if they do not.
279 // XXX - do f16 instructions?
280 class isFloatType<ValueType SrcVT> {
282 !if(!eq(SrcVT.Value, f16.Value), 1,
283 !if(!eq(SrcVT.Value, f32.Value), 1,
284 !if(!eq(SrcVT.Value, f64.Value), 1,
285 !if(!eq(SrcVT.Value, v2f16.Value), 1,
286 !if(!eq(SrcVT.Value, v4f16.Value), 1,
290 class isIntType<ValueType SrcVT> {
292 !if(!eq(SrcVT.Value, i16.Value), 1,
293 !if(!eq(SrcVT.Value, i32.Value), 1,
294 !if(!eq(SrcVT.Value, i64.Value), 1,
298 class isPackedType<ValueType SrcVT> {
300 !if(!eq(SrcVT.Value, v2i16.Value), 1,
301 !if(!eq(SrcVT.Value, v2f16.Value), 1,
302 !if(!eq(SrcVT.Value, v4f16.Value), 1, 0)
306 //===----------------------------------------------------------------------===//
307 // PatFrags for global memory operations
308 //===----------------------------------------------------------------------===//
310 foreach as = [ "global", "flat", "constant", "local", "private", "region" ] in {
311 let AddressSpaces = !cast<AddressSpaceList>("LoadAddress_"#as).AddrSpaces in {
314 defm atomic_inc_#as : binary_atomic_op<SIatomic_inc>;
315 defm atomic_dec_#as : binary_atomic_op<SIatomic_dec>;
316 defm atomic_load_fmin_#as : binary_atomic_op<SIatomic_fmin, 0>;
317 defm atomic_load_fmax_#as : binary_atomic_op<SIatomic_fmax, 0>;
320 } // End let AddressSpaces = ...
321 } // End foreach AddrSpace
323 def atomic_fadd_global_noret : PatFrag<
324 (ops node:$ptr, node:$value),
325 (SIglobal_atomic_fadd node:$ptr, node:$value)> {
329 let AddressSpaces = StoreAddress_global.AddrSpaces;
332 def atomic_pk_fadd_global_noret : PatFrag<
333 (ops node:$ptr, node:$value),
334 (SIglobal_atomic_pk_fadd node:$ptr, node:$value)> {
336 let MemoryVT = v2f16;
338 let AddressSpaces = StoreAddress_global.AddrSpaces;
341 //===----------------------------------------------------------------------===//
342 // SDNodes PatFrags for loads/stores with a glue input.
343 // This is for SDNodes and PatFrag for local loads and stores to
344 // enable s_mov_b32 m0, -1 to be glued to the memory instructions.
346 // These mirror the regular load/store PatFrags and rely on special
347 // processing during Select() to add the glued copy.
349 //===----------------------------------------------------------------------===//
351 def AMDGPUld_glue : SDNode <"ISD::LOAD", SDTLoad,
352 [SDNPHasChain, SDNPMayLoad, SDNPMemOperand, SDNPInGlue]
355 def AMDGPUatomic_ld_glue : SDNode <"ISD::ATOMIC_LOAD", SDTAtomicLoad,
356 [SDNPHasChain, SDNPMayLoad, SDNPMemOperand, SDNPInGlue]
359 def unindexedload_glue : PatFrag <(ops node:$ptr), (AMDGPUld_glue node:$ptr)> {
364 def load_glue : PatFrag <(ops node:$ptr), (unindexedload_glue node:$ptr)> {
366 let IsNonExtLoad = 1;
369 def atomic_load_32_glue : PatFrag<(ops node:$ptr),
370 (AMDGPUatomic_ld_glue node:$ptr)> {
375 def atomic_load_64_glue : PatFrag<(ops node:$ptr),
376 (AMDGPUatomic_ld_glue node:$ptr)> {
381 def extload_glue : PatFrag<(ops node:$ptr), (unindexedload_glue node:$ptr)> {
383 let IsAnyExtLoad = 1;
386 def sextload_glue : PatFrag<(ops node:$ptr), (unindexedload_glue node:$ptr)> {
388 let IsSignExtLoad = 1;
391 def zextload_glue : PatFrag<(ops node:$ptr), (unindexedload_glue node:$ptr)> {
393 let IsZeroExtLoad = 1;
396 def extloadi8_glue : PatFrag<(ops node:$ptr), (extload_glue node:$ptr)> {
401 def zextloadi8_glue : PatFrag<(ops node:$ptr), (zextload_glue node:$ptr)> {
406 def extloadi16_glue : PatFrag<(ops node:$ptr), (extload_glue node:$ptr)> {
411 def zextloadi16_glue : PatFrag<(ops node:$ptr), (zextload_glue node:$ptr)> {
416 def sextloadi8_glue : PatFrag<(ops node:$ptr), (sextload_glue node:$ptr)> {
421 def sextloadi16_glue : PatFrag<(ops node:$ptr), (sextload_glue node:$ptr)> {
427 let IsLoad = 1, AddressSpaces = LoadAddress_local.AddrSpaces in {
428 def load_local_m0 : PatFrag<(ops node:$ptr), (load_glue node:$ptr)> {
429 let IsNonExtLoad = 1;
432 let MemoryVT = i8 in {
433 def extloadi8_local_m0 : PatFrag<(ops node:$ptr), (extloadi8_glue node:$ptr)>;
434 def sextloadi8_local_m0 : PatFrag<(ops node:$ptr), (sextloadi8_glue node:$ptr)>;
435 def zextloadi8_local_m0 : PatFrag<(ops node:$ptr), (zextloadi8_glue node:$ptr)>;
438 let MemoryVT = i16 in {
439 def extloadi16_local_m0 : PatFrag<(ops node:$ptr), (extloadi16_glue node:$ptr)>;
440 def sextloadi16_local_m0 : PatFrag<(ops node:$ptr), (sextloadi16_glue node:$ptr)>;
441 def zextloadi16_local_m0 : PatFrag<(ops node:$ptr), (zextloadi16_glue node:$ptr)>;
444 def load_align8_local_m0 : PatFrag<(ops node:$ptr),
445 (load_local_m0 node:$ptr)> {
447 let IsNonExtLoad = 1;
448 let MinAlignment = 8;
450 def load_align16_local_m0 : PatFrag<(ops node:$ptr),
451 (load_local_m0 node:$ptr)> {
453 let IsNonExtLoad = 1;
454 let MinAlignment = 16;
459 let IsAtomic = 1, AddressSpaces = LoadAddress_local.AddrSpaces in {
460 def atomic_load_32_local_m0 : PatFrag<(ops node:$ptr),
461 (atomic_load_32_glue node:$ptr)> {
464 def atomic_load_64_local_m0 : PatFrag<(ops node:$ptr),
465 (atomic_load_64_glue node:$ptr)> {
469 } // End let AddressSpaces = LoadAddress_local.AddrSpaces
472 def AMDGPUst_glue : SDNode <"ISD::STORE", SDTStore,
473 [SDNPHasChain, SDNPMayStore, SDNPMemOperand, SDNPInGlue]
476 def AMDGPUatomic_st_glue : SDNode <"ISD::ATOMIC_STORE", SDTAtomicStore,
477 [SDNPHasChain, SDNPMayStore, SDNPMemOperand, SDNPInGlue]
480 def unindexedstore_glue : PatFrag<(ops node:$val, node:$ptr),
481 (AMDGPUst_glue node:$val, node:$ptr)> {
486 def store_glue : PatFrag<(ops node:$val, node:$ptr),
487 (unindexedstore_glue node:$val, node:$ptr)> {
489 let IsTruncStore = 0;
492 def truncstore_glue : PatFrag<(ops node:$val, node:$ptr),
493 (unindexedstore_glue node:$val, node:$ptr)> {
495 let IsTruncStore = 1;
498 def truncstorei8_glue : PatFrag<(ops node:$val, node:$ptr),
499 (truncstore_glue node:$val, node:$ptr)> {
504 def truncstorei16_glue : PatFrag<(ops node:$val, node:$ptr),
505 (truncstore_glue node:$val, node:$ptr)> {
510 let IsStore = 1, AddressSpaces = StoreAddress_local.AddrSpaces in {
511 def store_local_m0 : PatFrag<(ops node:$val, node:$ptr),
512 (store_glue node:$val, node:$ptr)> {
514 let IsTruncStore = 0;
517 def truncstorei8_local_m0 : PatFrag<(ops node:$val, node:$ptr),
518 (unindexedstore_glue node:$val, node:$ptr)> {
523 def truncstorei16_local_m0 : PatFrag<(ops node:$val, node:$ptr),
524 (unindexedstore_glue node:$val, node:$ptr)> {
530 def store_align16_local_m0 : PatFrag <
531 (ops node:$value, node:$ptr),
532 (store_local_m0 node:$value, node:$ptr)> {
534 let IsTruncStore = 0;
535 let MinAlignment = 16;
538 def store_align8_local_m0 : PatFrag <
539 (ops node:$value, node:$ptr),
540 (store_local_m0 node:$value, node:$ptr)> {
542 let IsTruncStore = 0;
543 let MinAlignment = 8;
546 let AddressSpaces = StoreAddress_local.AddrSpaces in {
548 def atomic_store_local_32_m0 : PatFrag <
549 (ops node:$value, node:$ptr),
550 (AMDGPUatomic_st_glue node:$value, node:$ptr)> {
554 def atomic_store_local_64_m0 : PatFrag <
555 (ops node:$value, node:$ptr),
556 (AMDGPUatomic_st_glue node:$value, node:$ptr)> {
560 } // End let AddressSpaces = StoreAddress_local.AddrSpaces
563 def si_setcc_uniform : PatFrag <
564 (ops node:$lhs, node:$rhs, node:$cond),
565 (setcc node:$lhs, node:$rhs, node:$cond), [{
566 for (SDNode *Use : N->uses()) {
567 if (Use->isMachineOpcode() || Use->getOpcode() != ISD::CopyToReg)
570 unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
571 if (Reg != AMDGPU::SCC)
577 //===----------------------------------------------------------------------===//
578 // SDNodes PatFrags for d16 loads
579 //===----------------------------------------------------------------------===//
581 class LoadD16Frag <SDPatternOperator op> : PatFrag<(ops node:$ptr, node:$tied_in), (op node:$ptr, node:$tied_in)>;
582 class LocalLoadD16 <SDPatternOperator op> : LoadD16Frag <op>, LocalAddress;
583 class GlobalLoadD16 <SDPatternOperator op> : LoadD16Frag <op>, GlobalLoadAddress;
584 class PrivateLoadD16 <SDPatternOperator op> : LoadD16Frag <op>, PrivateAddress;
585 class FlatLoadD16 <SDPatternOperator op> : LoadD16Frag <op>, FlatLoadAddress;
587 def load_d16_hi_local : LocalLoadD16 <SIload_d16_hi>;
588 def az_extloadi8_d16_hi_local : LocalLoadD16 <SIload_d16_hi_u8>;
589 def sextloadi8_d16_hi_local : LocalLoadD16 <SIload_d16_hi_i8>;
591 def load_d16_hi_global : GlobalLoadD16 <SIload_d16_hi>;
592 def az_extloadi8_d16_hi_global : GlobalLoadD16 <SIload_d16_hi_u8>;
593 def sextloadi8_d16_hi_global : GlobalLoadD16 <SIload_d16_hi_i8>;
595 def load_d16_hi_private : PrivateLoadD16 <SIload_d16_hi>;
596 def az_extloadi8_d16_hi_private : PrivateLoadD16 <SIload_d16_hi_u8>;
597 def sextloadi8_d16_hi_private : PrivateLoadD16 <SIload_d16_hi_i8>;
599 def load_d16_hi_flat : FlatLoadD16 <SIload_d16_hi>;
600 def az_extloadi8_d16_hi_flat : FlatLoadD16 <SIload_d16_hi_u8>;
601 def sextloadi8_d16_hi_flat : FlatLoadD16 <SIload_d16_hi_i8>;
604 def load_d16_lo_local : LocalLoadD16 <SIload_d16_lo>;
605 def az_extloadi8_d16_lo_local : LocalLoadD16 <SIload_d16_lo_u8>;
606 def sextloadi8_d16_lo_local : LocalLoadD16 <SIload_d16_lo_i8>;
608 def load_d16_lo_global : GlobalLoadD16 <SIload_d16_lo>;
609 def az_extloadi8_d16_lo_global : GlobalLoadD16 <SIload_d16_lo_u8>;
610 def sextloadi8_d16_lo_global : GlobalLoadD16 <SIload_d16_lo_i8>;
612 def load_d16_lo_private : PrivateLoadD16 <SIload_d16_lo>;
613 def az_extloadi8_d16_lo_private : PrivateLoadD16 <SIload_d16_lo_u8>;
614 def sextloadi8_d16_lo_private : PrivateLoadD16 <SIload_d16_lo_i8>;
616 def load_d16_lo_flat : FlatLoadD16 <SIload_d16_lo>;
617 def az_extloadi8_d16_lo_flat : FlatLoadD16 <SIload_d16_lo_u8>;
618 def sextloadi8_d16_lo_flat : FlatLoadD16 <SIload_d16_lo_i8>;
622 def lshr_rev : PatFrag <
623 (ops node:$src1, node:$src0),
627 def ashr_rev : PatFrag <
628 (ops node:$src1, node:$src0),
632 def lshl_rev : PatFrag <
633 (ops node:$src1, node:$src0),
637 def add_ctpop : PatFrag <
638 (ops node:$src0, node:$src1),
639 (add (ctpop $src0), $src1)
642 multiclass SIAtomicM0Glue2 <string op_name, bit is_amdgpu = 0,
643 SDTypeProfile tc = SDTAtomic2,
647 !if(is_amdgpu, "AMDGPUISD", "ISD")#"::ATOMIC_"#op_name, tc,
648 [SDNPHasChain, SDNPMayStore, SDNPMayLoad, SDNPMemOperand, SDNPInGlue]
651 let AddressSpaces = StoreAddress_local.AddrSpaces in {
652 defm _local_m0 : binary_atomic_op <!cast<SDNode>(NAME#"_glue"), IsInt>;
655 let AddressSpaces = StoreAddress_region.AddrSpaces in {
656 defm _region_m0 : binary_atomic_op <!cast<SDNode>(NAME#"_glue"), IsInt>;
660 defm atomic_load_add : SIAtomicM0Glue2 <"LOAD_ADD">;
661 defm atomic_load_sub : SIAtomicM0Glue2 <"LOAD_SUB">;
662 defm atomic_inc : SIAtomicM0Glue2 <"INC", 1>;
663 defm atomic_dec : SIAtomicM0Glue2 <"DEC", 1>;
664 defm atomic_load_and : SIAtomicM0Glue2 <"LOAD_AND">;
665 defm atomic_load_min : SIAtomicM0Glue2 <"LOAD_MIN">;
666 defm atomic_load_max : SIAtomicM0Glue2 <"LOAD_MAX">;
667 defm atomic_load_or : SIAtomicM0Glue2 <"LOAD_OR">;
668 defm atomic_load_xor : SIAtomicM0Glue2 <"LOAD_XOR">;
669 defm atomic_load_umin : SIAtomicM0Glue2 <"LOAD_UMIN">;
670 defm atomic_load_umax : SIAtomicM0Glue2 <"LOAD_UMAX">;
671 defm atomic_swap : SIAtomicM0Glue2 <"SWAP">;
672 defm atomic_load_fadd : SIAtomicM0Glue2 <"LOAD_FADD", 0, SDTAtomic2_f32, 0>;
673 defm atomic_load_fmin : SIAtomicM0Glue2 <"LOAD_FMIN", 1, SDTAtomic2_f32, 0>;
674 defm atomic_load_fmax : SIAtomicM0Glue2 <"LOAD_FMAX", 1, SDTAtomic2_f32, 0>;
676 def as_i1imm : SDNodeXForm<imm, [{
677 return CurDAG->getTargetConstant(N->getZExtValue(), SDLoc(N), MVT::i1);
680 def as_i8imm : SDNodeXForm<imm, [{
681 return CurDAG->getTargetConstant(N->getZExtValue(), SDLoc(N), MVT::i8);
684 def as_i16imm : SDNodeXForm<imm, [{
685 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i16);
688 def as_i32imm: SDNodeXForm<imm, [{
689 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32);
692 def as_i64imm: SDNodeXForm<imm, [{
693 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i64);
696 def cond_as_i32imm: SDNodeXForm<cond, [{
697 return CurDAG->getTargetConstant(N->get(), SDLoc(N), MVT::i32);
700 // Copied from the AArch64 backend:
701 def bitcast_fpimm_to_i32 : SDNodeXForm<fpimm, [{
702 return CurDAG->getTargetConstant(
703 N->getValueAPF().bitcastToAPInt().getZExtValue(), SDLoc(N), MVT::i32);
706 def frameindex_to_targetframeindex : SDNodeXForm<frameindex, [{
707 auto FI = cast<FrameIndexSDNode>(N);
708 return CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
711 // Copied from the AArch64 backend:
712 def bitcast_fpimm_to_i64 : SDNodeXForm<fpimm, [{
713 return CurDAG->getTargetConstant(
714 N->getValueAPF().bitcastToAPInt().getZExtValue(), SDLoc(N), MVT::i64);
717 class bitextract_imm<int bitnum> : SDNodeXForm<imm, [{
718 uint64_t Imm = N->getZExtValue();
719 unsigned Bit = (Imm >> }] # bitnum # [{ ) & 1;
720 return CurDAG->getTargetConstant(Bit, SDLoc(N), MVT::i1);
723 def SIMM16bit : ImmLeaf <i32,
724 [{return isInt<16>(Imm);}]
727 def UIMM16bit : ImmLeaf <i32,
728 [{return isUInt<16>(Imm);}]
731 class InlineImm <ValueType vt> : PatLeaf <(vt imm), [{
732 return isInlineImmediate(N);
735 class InlineFPImm <ValueType vt> : PatLeaf <(vt fpimm), [{
736 return isInlineImmediate(N);
739 class VGPRImm <dag frag> : PatLeaf<frag, [{
743 def NegateImm : SDNodeXForm<imm, [{
744 return CurDAG->getConstant(-N->getSExtValue(), SDLoc(N), MVT::i32);
747 // TODO: When FP inline imm values work?
748 def NegSubInlineConst32 : ImmLeaf<i32, [{
749 return Imm < -16 && Imm >= -64;
752 def NegSubInlineConst16 : ImmLeaf<i16, [{
753 return Imm < -16 && Imm >= -64;
756 def ShiftAmt32Imm : PatLeaf <(imm), [{
757 return N->getZExtValue() < 32;
760 def getNegV2I16Imm : SDNodeXForm<build_vector, [{
761 return SDValue(packNegConstantV2I16(N, *CurDAG), 0);
764 def NegSubInlineConstV216 : PatLeaf<(build_vector), [{
765 assert(N->getNumOperands() == 2);
766 assert(N->getOperand(0).getValueType().getSizeInBits() == 16);
767 SDValue Src0 = N->getOperand(0);
768 SDValue Src1 = N->getOperand(1);
770 return isNegInlineImmediate(Src0.getNode());
772 return (isNullConstantOrUndef(Src0) && isNegInlineImmediate(Src1.getNode())) ||
773 (isNullConstantOrUndef(Src1) && isNegInlineImmediate(Src0.getNode()));
776 //===----------------------------------------------------------------------===//
778 //===----------------------------------------------------------------------===//
780 def SoppBrTarget : AsmOperandClass {
781 let Name = "SoppBrTarget";
782 let ParserMethod = "parseSOppBrTarget";
785 def sopp_brtarget : Operand<OtherVT> {
786 let EncoderMethod = "getSOPPBrEncoding";
787 let DecoderMethod = "decodeSoppBrTarget";
788 let OperandType = "OPERAND_PCREL";
789 let ParserMatchClass = SoppBrTarget;
792 def si_ga : Operand<iPTR>;
794 def InterpSlotMatchClass : AsmOperandClass {
795 let Name = "InterpSlot";
796 let PredicateMethod = "isInterpSlot";
797 let ParserMethod = "parseInterpSlot";
798 let RenderMethod = "addImmOperands";
801 def InterpSlot : Operand<i32> {
802 let PrintMethod = "printInterpSlot";
803 let ParserMatchClass = InterpSlotMatchClass;
804 let OperandType = "OPERAND_IMMEDIATE";
807 def AttrMatchClass : AsmOperandClass {
809 let PredicateMethod = "isInterpAttr";
810 let ParserMethod = "parseInterpAttr";
811 let RenderMethod = "addImmOperands";
814 // It appears to be necessary to create a separate operand for this to
815 // be able to parse attr<num> with no space.
816 def Attr : Operand<i32> {
817 let PrintMethod = "printInterpAttr";
818 let ParserMatchClass = AttrMatchClass;
819 let OperandType = "OPERAND_IMMEDIATE";
822 def AttrChanMatchClass : AsmOperandClass {
823 let Name = "AttrChan";
824 let PredicateMethod = "isAttrChan";
825 let RenderMethod = "addImmOperands";
828 def AttrChan : Operand<i32> {
829 let PrintMethod = "printInterpAttrChan";
830 let ParserMatchClass = AttrChanMatchClass;
831 let OperandType = "OPERAND_IMMEDIATE";
834 def SendMsgMatchClass : AsmOperandClass {
835 let Name = "SendMsg";
836 let PredicateMethod = "isSendMsg";
837 let ParserMethod = "parseSendMsgOp";
838 let RenderMethod = "addImmOperands";
841 def SwizzleMatchClass : AsmOperandClass {
842 let Name = "Swizzle";
843 let PredicateMethod = "isSwizzle";
844 let ParserMethod = "parseSwizzleOp";
845 let RenderMethod = "addImmOperands";
849 def EndpgmMatchClass : AsmOperandClass {
850 let Name = "EndpgmImm";
851 let PredicateMethod = "isEndpgm";
852 let ParserMethod = "parseEndpgmOp";
853 let RenderMethod = "addImmOperands";
857 def ExpTgtMatchClass : AsmOperandClass {
859 let PredicateMethod = "isExpTgt";
860 let ParserMethod = "parseExpTgt";
861 let RenderMethod = "printExpTgt";
864 def SendMsgImm : Operand<i32> {
865 let PrintMethod = "printSendMsg";
866 let ParserMatchClass = SendMsgMatchClass;
869 def SwizzleImm : Operand<i16> {
870 let PrintMethod = "printSwizzle";
871 let ParserMatchClass = SwizzleMatchClass;
874 def EndpgmImm : Operand<i16> {
875 let PrintMethod = "printEndpgm";
876 let ParserMatchClass = EndpgmMatchClass;
879 def SWaitMatchClass : AsmOperandClass {
880 let Name = "SWaitCnt";
881 let RenderMethod = "addImmOperands";
882 let ParserMethod = "parseSWaitCntOps";
885 def VReg32OrOffClass : AsmOperandClass {
886 let Name = "VReg32OrOff";
887 let ParserMethod = "parseVReg32OrOff";
890 def WAIT_FLAG : Operand <i32> {
891 let ParserMatchClass = SWaitMatchClass;
892 let PrintMethod = "printWaitFlag";
893 let OperandType = "OPERAND_IMMEDIATE";
896 include "SIInstrFormats.td"
897 include "VIInstrFormats.td"
899 def BoolReg : AsmOperandClass {
900 let Name = "BoolReg";
901 let ParserMethod = "parseBoolReg";
902 let RenderMethod = "addRegOperands";
905 class BoolRC : RegisterOperand<SReg_1> {
906 let ParserMatchClass = BoolReg;
907 let DecoderMethod = "decodeBoolReg";
910 def SSrc_i1 : RegisterOperand<SReg_1_XEXEC> {
911 let ParserMatchClass = BoolReg;
912 let DecoderMethod = "decodeBoolReg";
915 def VOPDstS64orS32 : BoolRC {
916 let PrintMethod = "printVOPDst";
919 // SCSrc_i1 is the operand for pseudo instructions only.
920 // Boolean immeadiates shall not be exposed to codegen instructions.
921 def SCSrc_i1 : RegisterOperand<SReg_1_XEXEC> {
922 let OperandNamespace = "AMDGPU";
923 let OperandType = "OPERAND_REG_IMM_INT32";
924 let ParserMatchClass = BoolReg;
925 let DecoderMethod = "decodeBoolReg";
928 // ===----------------------------------------------------------------------===//
929 // ExpSrc* Special cases for exp src operands which are printed as
930 // "off" depending on en operand.
931 // ===----------------------------------------------------------------------===//
933 def ExpSrc0 : RegisterOperand<VGPR_32> {
934 let PrintMethod = "printExpSrc0";
935 let ParserMatchClass = VReg32OrOffClass;
938 def ExpSrc1 : RegisterOperand<VGPR_32> {
939 let PrintMethod = "printExpSrc1";
940 let ParserMatchClass = VReg32OrOffClass;
943 def ExpSrc2 : RegisterOperand<VGPR_32> {
944 let PrintMethod = "printExpSrc2";
945 let ParserMatchClass = VReg32OrOffClass;
948 def ExpSrc3 : RegisterOperand<VGPR_32> {
949 let PrintMethod = "printExpSrc3";
950 let ParserMatchClass = VReg32OrOffClass;
953 class SDWASrc<ValueType vt> : RegisterOperand<VS_32> {
954 let OperandNamespace = "AMDGPU";
955 string Type = !if(isFloatType<vt>.ret, "FP", "INT");
956 let OperandType = "OPERAND_REG_INLINE_C_"#Type#vt.Size;
957 let DecoderMethod = "decodeSDWASrc"#vt.Size;
958 let EncoderMethod = "getSDWASrcEncoding";
961 def SDWASrc_i32 : SDWASrc<i32>;
962 def SDWASrc_i16 : SDWASrc<i16>;
963 def SDWASrc_f32 : SDWASrc<f32>;
964 def SDWASrc_f16 : SDWASrc<f16>;
966 def SDWAVopcDst : BoolRC {
967 let OperandNamespace = "AMDGPU";
968 let OperandType = "OPERAND_SDWA_VOPC_DST";
969 let EncoderMethod = "getSDWAVopcDstEncoding";
970 let DecoderMethod = "decodeSDWAVopcDst";
971 let PrintMethod = "printVOPDst";
974 class NamedMatchClass<string CName, bit Optional = 1> : AsmOperandClass {
975 let Name = "Imm"#CName;
976 let PredicateMethod = "is"#CName;
977 let ParserMethod = !if(Optional, "parseOptionalOperand", "parse"#CName);
978 let RenderMethod = "addImmOperands";
979 let IsOptional = Optional;
980 let DefaultMethod = !if(Optional, "default"#CName, ?);
983 class NamedOperandBit<string Name, AsmOperandClass MatchClass> : Operand<i1> {
984 let PrintMethod = "print"#Name;
985 let ParserMatchClass = MatchClass;
988 class NamedOperandU8<string Name, AsmOperandClass MatchClass> : Operand<i8> {
989 let PrintMethod = "print"#Name;
990 let ParserMatchClass = MatchClass;
993 class NamedOperandU16<string Name, AsmOperandClass MatchClass> : Operand<i16> {
994 let PrintMethod = "print"#Name;
995 let ParserMatchClass = MatchClass;
998 class NamedOperandU32<string Name, AsmOperandClass MatchClass> : Operand<i32> {
999 let PrintMethod = "print"#Name;
1000 let ParserMatchClass = MatchClass;
1003 class NamedOperandU32Default0<string Name, AsmOperandClass MatchClass> :
1004 OperandWithDefaultOps<i32, (ops (i32 0))> {
1005 let PrintMethod = "print"#Name;
1006 let ParserMatchClass = MatchClass;
1009 let OperandType = "OPERAND_IMMEDIATE" in {
1011 def offen : NamedOperandBit<"Offen", NamedMatchClass<"Offen">>;
1012 def idxen : NamedOperandBit<"Idxen", NamedMatchClass<"Idxen">>;
1013 def addr64 : NamedOperandBit<"Addr64", NamedMatchClass<"Addr64">>;
1015 def flat_offset : NamedOperandU16<"FlatOffset", NamedMatchClass<"FlatOffset">>;
1016 def offset : NamedOperandU16<"Offset", NamedMatchClass<"Offset">>;
1017 def offset0 : NamedOperandU8<"Offset0", NamedMatchClass<"Offset0">>;
1018 def offset1 : NamedOperandU8<"Offset1", NamedMatchClass<"Offset1">>;
1020 def gds : NamedOperandBit<"GDS", NamedMatchClass<"GDS">>;
1022 def omod : NamedOperandU32<"OModSI", NamedMatchClass<"OModSI">>;
1023 def clampmod : NamedOperandBit<"ClampSI", NamedMatchClass<"ClampSI">>;
1024 def highmod : NamedOperandBit<"High", NamedMatchClass<"High">>;
1026 def DLC : NamedOperandBit<"DLC", NamedMatchClass<"DLC">>;
1027 def GLC : NamedOperandBit<"GLC", NamedMatchClass<"GLC">>;
1028 def SLC : NamedOperandBit<"SLC", NamedMatchClass<"SLC">>;
1029 def TFE : NamedOperandBit<"TFE", NamedMatchClass<"TFE">>;
1030 def UNorm : NamedOperandBit<"UNorm", NamedMatchClass<"UNorm">>;
1031 def DA : NamedOperandBit<"DA", NamedMatchClass<"DA">>;
1032 def R128A16 : NamedOperandBit<"R128A16", NamedMatchClass<"R128A16">>;
1033 def D16 : NamedOperandBit<"D16", NamedMatchClass<"D16">>;
1034 def LWE : NamedOperandBit<"LWE", NamedMatchClass<"LWE">>;
1035 def exp_compr : NamedOperandBit<"ExpCompr", NamedMatchClass<"ExpCompr">>;
1036 def exp_vm : NamedOperandBit<"ExpVM", NamedMatchClass<"ExpVM">>;
1038 def FORMAT : NamedOperandU8<"FORMAT", NamedMatchClass<"FORMAT">>;
1040 def DMask : NamedOperandU16<"DMask", NamedMatchClass<"DMask">>;
1041 def Dim : NamedOperandU8<"Dim", NamedMatchClass<"Dim", 0>>;
1043 def dpp8 : NamedOperandU32<"DPP8", NamedMatchClass<"DPP8", 0>>;
1045 def dpp_ctrl : NamedOperandU32<"DPPCtrl", NamedMatchClass<"DPPCtrl", 0>>;
1046 def row_mask : NamedOperandU32<"RowMask", NamedMatchClass<"RowMask">>;
1047 def bank_mask : NamedOperandU32<"BankMask", NamedMatchClass<"BankMask">>;
1048 def bound_ctrl : NamedOperandBit<"BoundCtrl", NamedMatchClass<"BoundCtrl">>;
1049 def FI : NamedOperandU32<"FI", NamedMatchClass<"FI">>;
1051 def dst_sel : NamedOperandU32<"SDWADstSel", NamedMatchClass<"SDWADstSel">>;
1052 def src0_sel : NamedOperandU32<"SDWASrc0Sel", NamedMatchClass<"SDWASrc0Sel">>;
1053 def src1_sel : NamedOperandU32<"SDWASrc1Sel", NamedMatchClass<"SDWASrc1Sel">>;
1054 def dst_unused : NamedOperandU32<"SDWADstUnused", NamedMatchClass<"SDWADstUnused">>;
1056 def op_sel : NamedOperandU32Default0<"OpSel", NamedMatchClass<"OpSel">>;
1057 def op_sel_hi : NamedOperandU32Default0<"OpSelHi", NamedMatchClass<"OpSelHi">>;
1058 def neg_lo : NamedOperandU32Default0<"NegLo", NamedMatchClass<"NegLo">>;
1059 def neg_hi : NamedOperandU32Default0<"NegHi", NamedMatchClass<"NegHi">>;
1061 def blgp : NamedOperandU32<"BLGP", NamedMatchClass<"BLGP">>;
1062 def cbsz : NamedOperandU32<"CBSZ", NamedMatchClass<"CBSZ">>;
1063 def abid : NamedOperandU32<"ABID", NamedMatchClass<"ABID">>;
1065 def hwreg : NamedOperandU16<"Hwreg", NamedMatchClass<"Hwreg", 0>>;
1067 def exp_tgt : NamedOperandU8<"ExpTgt", NamedMatchClass<"ExpTgt", 0>> {
1071 } // End OperandType = "OPERAND_IMMEDIATE"
1073 class KImmMatchClass<int size> : AsmOperandClass {
1074 let Name = "KImmFP"#size;
1075 let PredicateMethod = "isKImmFP"#size;
1076 let ParserMethod = "parseImm";
1077 let RenderMethod = "addKImmFP"#size#"Operands";
1080 class kimmOperand<ValueType vt> : Operand<vt> {
1081 let OperandNamespace = "AMDGPU";
1082 let OperandType = "OPERAND_KIMM"#vt.Size;
1083 let PrintMethod = "printU"#vt.Size#"ImmOperand";
1084 let ParserMatchClass = !cast<AsmOperandClass>("KImmFP"#vt.Size#"MatchClass");
1087 // 32-bit VALU immediate operand that uses the constant bus.
1088 def KImmFP32MatchClass : KImmMatchClass<32>;
1089 def f32kimm : kimmOperand<i32>;
1091 // 32-bit VALU immediate operand with a 16-bit value that uses the
1093 def KImmFP16MatchClass : KImmMatchClass<16>;
1094 def f16kimm : kimmOperand<i16>;
1096 class FPInputModsMatchClass <int opSize> : AsmOperandClass {
1097 let Name = "RegOrImmWithFP"#opSize#"InputMods";
1098 let ParserMethod = "parseRegOrImmWithFPInputMods";
1099 let PredicateMethod = "isRegOrImmWithFP"#opSize#"InputMods";
1102 def FP16InputModsMatchClass : FPInputModsMatchClass<16>;
1103 def FP32InputModsMatchClass : FPInputModsMatchClass<32>;
1104 def FP64InputModsMatchClass : FPInputModsMatchClass<64>;
1106 class InputMods <AsmOperandClass matchClass> : Operand <i32> {
1107 let OperandNamespace = "AMDGPU";
1108 let OperandType = "OPERAND_INPUT_MODS";
1109 let ParserMatchClass = matchClass;
1112 class FPInputMods <FPInputModsMatchClass matchClass> : InputMods <matchClass> {
1113 let PrintMethod = "printOperandAndFPInputMods";
1116 def FP16InputMods : FPInputMods<FP16InputModsMatchClass>;
1117 def FP32InputMods : FPInputMods<FP32InputModsMatchClass>;
1118 def FP64InputMods : FPInputMods<FP64InputModsMatchClass>;
1120 class IntInputModsMatchClass <int opSize> : AsmOperandClass {
1121 let Name = "RegOrImmWithInt"#opSize#"InputMods";
1122 let ParserMethod = "parseRegOrImmWithIntInputMods";
1123 let PredicateMethod = "isRegOrImmWithInt"#opSize#"InputMods";
1125 def Int32InputModsMatchClass : IntInputModsMatchClass<32>;
1126 def Int64InputModsMatchClass : IntInputModsMatchClass<64>;
1128 class IntInputMods <IntInputModsMatchClass matchClass> : InputMods <matchClass> {
1129 let PrintMethod = "printOperandAndIntInputMods";
1131 def Int32InputMods : IntInputMods<Int32InputModsMatchClass>;
1132 def Int64InputMods : IntInputMods<Int64InputModsMatchClass>;
1134 class OpSelModsMatchClass : AsmOperandClass {
1135 let Name = "OpSelMods";
1136 let ParserMethod = "parseRegOrImm";
1137 let PredicateMethod = "isRegOrImm";
1140 def IntOpSelModsMatchClass : OpSelModsMatchClass;
1141 def IntOpSelMods : InputMods<IntOpSelModsMatchClass>;
1143 class FPSDWAInputModsMatchClass <int opSize> : AsmOperandClass {
1144 let Name = "SDWAWithFP"#opSize#"InputMods";
1145 let ParserMethod = "parseRegOrImmWithFPInputMods";
1146 let PredicateMethod = "isSDWAFP"#opSize#"Operand";
1149 def FP16SDWAInputModsMatchClass : FPSDWAInputModsMatchClass<16>;
1150 def FP32SDWAInputModsMatchClass : FPSDWAInputModsMatchClass<32>;
1152 class FPSDWAInputMods <FPSDWAInputModsMatchClass matchClass> :
1153 InputMods <matchClass> {
1154 let PrintMethod = "printOperandAndFPInputMods";
1157 def FP16SDWAInputMods : FPSDWAInputMods<FP16SDWAInputModsMatchClass>;
1158 def FP32SDWAInputMods : FPSDWAInputMods<FP32SDWAInputModsMatchClass>;
1160 def FPVRegInputModsMatchClass : AsmOperandClass {
1161 let Name = "VRegWithFPInputMods";
1162 let ParserMethod = "parseRegWithFPInputMods";
1163 let PredicateMethod = "isVReg32";
1166 def FPVRegInputMods : InputMods <FPVRegInputModsMatchClass> {
1167 let PrintMethod = "printOperandAndFPInputMods";
1170 class IntSDWAInputModsMatchClass <int opSize> : AsmOperandClass {
1171 let Name = "SDWAWithInt"#opSize#"InputMods";
1172 let ParserMethod = "parseRegOrImmWithIntInputMods";
1173 let PredicateMethod = "isSDWAInt"#opSize#"Operand";
1176 def Int16SDWAInputModsMatchClass : IntSDWAInputModsMatchClass<16>;
1177 def Int32SDWAInputModsMatchClass : IntSDWAInputModsMatchClass<32>;
1179 class IntSDWAInputMods <IntSDWAInputModsMatchClass matchClass> :
1180 InputMods <matchClass> {
1181 let PrintMethod = "printOperandAndIntInputMods";
1184 def Int16SDWAInputMods : IntSDWAInputMods<Int16SDWAInputModsMatchClass>;
1185 def Int32SDWAInputMods : IntSDWAInputMods<Int32SDWAInputModsMatchClass>;
1187 def IntVRegInputModsMatchClass : AsmOperandClass {
1188 let Name = "VRegWithIntInputMods";
1189 let ParserMethod = "parseRegWithIntInputMods";
1190 let PredicateMethod = "isVReg32";
1193 def IntVRegInputMods : InputMods <IntVRegInputModsMatchClass> {
1194 let PrintMethod = "printOperandAndIntInputMods";
1197 class PackedFPInputModsMatchClass <int opSize> : AsmOperandClass {
1198 let Name = "PackedFP"#opSize#"InputMods";
1199 let ParserMethod = "parseRegOrImm";
1200 let PredicateMethod = "isRegOrImm";
1201 // let PredicateMethod = "isPackedFP"#opSize#"InputMods";
1204 class PackedIntInputModsMatchClass <int opSize> : AsmOperandClass {
1205 let Name = "PackedInt"#opSize#"InputMods";
1206 let ParserMethod = "parseRegOrImm";
1207 let PredicateMethod = "isRegOrImm";
1208 // let PredicateMethod = "isPackedInt"#opSize#"InputMods";
1211 def PackedF16InputModsMatchClass : PackedFPInputModsMatchClass<16>;
1212 def PackedI16InputModsMatchClass : PackedIntInputModsMatchClass<16>;
1214 class PackedFPInputMods <PackedFPInputModsMatchClass matchClass> : InputMods <matchClass> {
1215 // let PrintMethod = "printPackedFPInputMods";
1218 class PackedIntInputMods <PackedIntInputModsMatchClass matchClass> : InputMods <matchClass> {
1219 //let PrintMethod = "printPackedIntInputMods";
1222 def PackedF16InputMods : PackedFPInputMods<PackedF16InputModsMatchClass>;
1223 def PackedI16InputMods : PackedIntInputMods<PackedI16InputModsMatchClass>;
1225 //===----------------------------------------------------------------------===//
1227 //===----------------------------------------------------------------------===//
1229 def DS1Addr1Offset : ComplexPattern<i32, 2, "SelectDS1Addr1Offset">;
1230 def DS64Bit4ByteAligned : ComplexPattern<i32, 3, "SelectDS64Bit4ByteAligned">;
1232 def MOVRELOffset : ComplexPattern<i32, 2, "SelectMOVRELOffset">;
1234 def VOP3Mods0 : ComplexPattern<untyped, 4, "SelectVOP3Mods0">;
1235 def VOP3Mods0Clamp : ComplexPattern<untyped, 3, "SelectVOP3Mods0Clamp">;
1236 def VOP3Mods0Clamp0OMod : ComplexPattern<untyped, 4, "SelectVOP3Mods0Clamp0OMod">;
1237 def VOP3Mods : ComplexPattern<untyped, 2, "SelectVOP3Mods">;
1238 def VOP3NoMods : ComplexPattern<untyped, 1, "SelectVOP3NoMods">;
1239 // VOP3Mods, but the input source is known to never be NaN.
1240 def VOP3Mods_nnan : ComplexPattern<fAny, 2, "SelectVOP3Mods_NNaN">;
1241 // VOP3Mods, but only allowed for f32 operands.
1242 def VOP3Mods_f32 : ComplexPattern<fAny, 2, "SelectVOP3Mods_f32">;
1244 def VOP3OMods : ComplexPattern<untyped, 3, "SelectVOP3OMods">;
1246 def VOP3PMods : ComplexPattern<untyped, 2, "SelectVOP3PMods">;
1247 def VOP3PMods0 : ComplexPattern<untyped, 3, "SelectVOP3PMods0">;
1249 def VOP3OpSel : ComplexPattern<untyped, 2, "SelectVOP3OpSel">;
1250 def VOP3OpSel0 : ComplexPattern<untyped, 3, "SelectVOP3OpSel0">;
1252 def VOP3OpSelMods : ComplexPattern<untyped, 2, "SelectVOP3OpSelMods">;
1253 def VOP3OpSelMods0 : ComplexPattern<untyped, 3, "SelectVOP3OpSelMods0">;
1255 def VOP3PMadMixMods : ComplexPattern<untyped, 2, "SelectVOP3PMadMixMods">;
1258 def Hi16Elt : ComplexPattern<untyped, 1, "SelectHi16Elt">;
1260 //===----------------------------------------------------------------------===//
1261 // SI assembler operands
1262 //===----------------------------------------------------------------------===//
1267 int FLAT_SCR = 0x68;
1270 // This should be kept in sync with SISrcMods enum
1294 int LLVM_DEBUG_TRAP = 3;
1310 int FLAT_SCR_LO = 20;
1311 int FLAT_SCR_HI = 21;
1312 int XNACK_MASK = 22;
1313 int POPS_PACKER = 25;
1316 class getHwRegImm<int Reg, int Offset = 0, int Size = 32> {
1318 !or(!shl(Offset, 6),
1319 !shl(!add(Size, -1), 11)));
1322 //===----------------------------------------------------------------------===//
1324 // SI Instruction multiclass helpers.
1326 // Instructions with _32 take 32-bit operands.
1327 // Instructions with _64 take 64-bit operands.
1329 // VOP_* instructions can use either a 32-bit or 64-bit encoding. The 32-bit
1330 // encoding is the standard encoding, but instruction that make use of
1331 // any of the instruction modifiers must use the 64-bit encoding.
1333 // Instructions with _e32 use the 32-bit encoding.
1334 // Instructions with _e64 use the 64-bit encoding.
1336 //===----------------------------------------------------------------------===//
1338 class SIMCInstr <string pseudo, int subtarget> {
1339 string PseudoInstr = pseudo;
1340 int Subtarget = subtarget;
1343 //===----------------------------------------------------------------------===//
1345 //===----------------------------------------------------------------------===//
1347 class EXP_Helper<bit done, SDPatternOperator node = null_frag> : EXPCommon<
1350 ExpSrc0:$src0, ExpSrc1:$src1, ExpSrc2:$src2, ExpSrc3:$src3,
1351 exp_vm:$vm, exp_compr:$compr, i8imm:$en),
1352 "exp$tgt $src0, $src1, $src2, $src3"#!if(done, " done", "")#"$compr$vm",
1353 [(node (i8 timm:$tgt), (i8 timm:$en),
1354 f32:$src0, f32:$src1, f32:$src2, f32:$src3,
1355 (i1 timm:$compr), (i1 timm:$vm))]> {
1356 let AsmMatchConverter = "cvtExp";
1359 // Split EXP instruction into EXP and EXP_DONE so we can set
1360 // mayLoad for done=1.
1361 multiclass EXP_m<bit done, SDPatternOperator node> {
1362 let mayLoad = done, DisableWQM = 1 in {
1363 let isPseudo = 1, isCodeGenOnly = 1 in {
1364 def "" : EXP_Helper<done, node>,
1365 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.NONE>;
1368 let done = done in {
1369 def _si : EXP_Helper<done>,
1370 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.SI>,
1372 let AssemblerPredicates = [isGFX6GFX7];
1373 let DecoderNamespace = "GFX6GFX7";
1374 let DisableDecoder = DisableSIDecoder;
1377 def _vi : EXP_Helper<done>,
1378 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.VI>,
1380 let AssemblerPredicates = [isGFX8GFX9];
1381 let DecoderNamespace = "GFX8";
1382 let DisableDecoder = DisableVIDecoder;
1385 def _gfx10 : EXP_Helper<done>,
1386 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.GFX10>,
1388 let AssemblerPredicates = [isGFX10Plus];
1389 let DecoderNamespace = "GFX10";
1390 let DisableDecoder = DisableSIDecoder;
1396 //===----------------------------------------------------------------------===//
1397 // Vector ALU classes
1398 //===----------------------------------------------------------------------===//
1400 class getNumSrcArgs<ValueType Src0, ValueType Src1, ValueType Src2> {
1402 !if (!eq(Src0.Value, untyped.Value), 0,
1403 !if (!eq(Src1.Value, untyped.Value), 1, // VOP1
1404 !if (!eq(Src2.Value, untyped.Value), 2, // VOP2
1408 // Returns the register class to use for the destination of VOP[123C]
1409 // instructions for the given VT.
1410 class getVALUDstForVT<ValueType VT> {
1411 RegisterOperand ret = !if(!eq(VT.Size, 32), VOPDstOperand<VGPR_32>,
1412 !if(!eq(VT.Size, 128), VOPDstOperand<VReg_128>,
1413 !if(!eq(VT.Size, 64), VOPDstOperand<VReg_64>,
1414 !if(!eq(VT.Size, 16), VOPDstOperand<VGPR_32>,
1415 VOPDstS64orS32)))); // else VT == i1
1418 // Returns true if VT is floating point.
1419 class getIsFP<ValueType VT> {
1420 bit ret = !if(!eq(VT.Value, f16.Value), 1,
1421 !if(!eq(VT.Value, v2f16.Value), 1,
1422 !if(!eq(VT.Value, v4f16.Value), 1,
1423 !if(!eq(VT.Value, f32.Value), 1,
1424 !if(!eq(VT.Value, v2f32.Value), 1,
1425 !if(!eq(VT.Value, f64.Value), 1,
1426 !if(!eq(VT.Value, v2f64.Value), 1,
1430 // Returns the register class to use for the destination of VOP[12C]
1431 // instructions with SDWA extension
1432 class getSDWADstForVT<ValueType VT> {
1433 RegisterOperand ret = !if(!eq(VT.Size, 1),
1434 SDWAVopcDst, // VOPC
1435 VOPDstOperand<VGPR_32>); // VOP1/2 32-bit dst
1438 // Returns the register class to use for source 0 of VOP[12C]
1439 // instructions for the given VT.
1440 class getVOPSrc0ForVT<ValueType VT> {
1441 bit isFP = getIsFP<VT>.ret;
1443 RegisterOperand ret =
1445 !if(!eq(VT.Size, 64),
1447 !if(!eq(VT.Value, f16.Value),
1449 !if(!eq(VT.Value, v2f16.Value),
1451 !if(!eq(VT.Value, v4f16.Value),
1458 !if(!eq(VT.Size, 64),
1460 !if(!eq(VT.Value, i16.Value),
1462 !if(!eq(VT.Value, v2i16.Value),
1471 // Returns the vreg register class to use for source operand given VT
1472 class getVregSrcForVT<ValueType VT> {
1473 RegisterClass ret = !if(!eq(VT.Size, 128), VReg_128,
1474 !if(!eq(VT.Size, 96), VReg_96,
1475 !if(!eq(VT.Size, 64), VReg_64,
1476 !if(!eq(VT.Size, 48), VReg_64,
1480 class getSDWASrcForVT <ValueType VT> {
1481 bit isFP = getIsFP<VT>.ret;
1482 RegisterOperand retFlt = !if(!eq(VT.Size, 16), SDWASrc_f16, SDWASrc_f32);
1483 RegisterOperand retInt = !if(!eq(VT.Size, 16), SDWASrc_i16, SDWASrc_i32);
1484 RegisterOperand ret = !if(isFP, retFlt, retInt);
1487 // Returns the register class to use for sources of VOP3 instructions for the
1489 class getVOP3SrcForVT<ValueType VT> {
1490 bit isFP = getIsFP<VT>.ret;
1491 RegisterOperand ret =
1492 !if(!eq(VT.Size, 128),
1494 !if(!eq(VT.Size, 64),
1498 !if(!eq(VT.Value, i1.Value),
1501 !if(!eq(VT.Value, f16.Value),
1503 !if(!eq(VT.Value, v2f16.Value),
1505 !if(!eq(VT.Value, v4f16.Value),
1511 !if(!eq(VT.Value, i16.Value),
1513 !if(!eq(VT.Value, v2i16.Value),
1524 // Float or packed int
1525 class isModifierType<ValueType SrcVT> {
1527 !if(!eq(SrcVT.Value, f16.Value), 1,
1528 !if(!eq(SrcVT.Value, f32.Value), 1,
1529 !if(!eq(SrcVT.Value, f64.Value), 1,
1530 !if(!eq(SrcVT.Value, v2f16.Value), 1,
1531 !if(!eq(SrcVT.Value, v2i16.Value), 1,
1535 // Return type of input modifiers operand for specified input operand
1536 class getSrcMod <ValueType VT, bit EnableF32SrcMods> {
1537 bit isFP = getIsFP<VT>.ret;
1538 bit isPacked = isPackedType<VT>.ret;
1539 Operand ret = !if(!eq(VT.Size, 64),
1540 !if(isFP, FP64InputMods, Int64InputMods),
1542 !if(!eq(VT.Value, f16.Value),
1546 !if(EnableF32SrcMods, FP32InputMods, Int32InputMods))
1550 class getOpSelMod <ValueType VT> {
1551 Operand ret = !if(!eq(VT.Value, f16.Value), FP16InputMods, IntOpSelMods);
1554 // Return type of input modifiers operand specified input operand for DPP
1555 class getSrcModExt <ValueType VT> {
1556 bit isFP = getIsFP<VT>.ret;
1557 Operand ret = !if(isFP, FPVRegInputMods, IntVRegInputMods);
1560 // Return type of input modifiers operand specified input operand for SDWA
1561 class getSrcModSDWA <ValueType VT> {
1562 Operand ret = !if(!eq(VT.Value, f16.Value), FP16SDWAInputMods,
1563 !if(!eq(VT.Value, f32.Value), FP32SDWAInputMods,
1564 !if(!eq(VT.Value, i16.Value), Int16SDWAInputMods,
1565 Int32SDWAInputMods)));
1568 // Returns the input arguments for VOP[12C] instructions for the given SrcVT.
1569 class getIns32 <RegisterOperand Src0RC, RegisterClass Src1RC, int NumSrcArgs> {
1570 dag ret = !if(!eq(NumSrcArgs, 1), (ins Src0RC:$src0), // VOP1
1571 !if(!eq(NumSrcArgs, 2), (ins Src0RC:$src0, Src1RC:$src1), // VOP2
1575 // Returns the input arguments for VOP3 instructions for the given SrcVT.
1576 class getIns64 <RegisterOperand Src0RC, RegisterOperand Src1RC,
1577 RegisterOperand Src2RC, int NumSrcArgs,
1578 bit HasIntClamp, bit HasModifiers, bit HasSrc2Mods, bit HasOMod,
1579 Operand Src0Mod, Operand Src1Mod, Operand Src2Mod> {
1582 !if (!eq(NumSrcArgs, 0),
1583 // VOP1 without input operands (V_NOP, V_CLREXCP)
1586 !if (!eq(NumSrcArgs, 1),
1587 !if (!eq(HasModifiers, 1),
1588 // VOP1 with modifiers
1589 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1590 clampmod:$clamp, omod:$omod)
1592 // VOP1 without modifiers
1593 !if (!eq(HasIntClamp, 1),
1594 (ins Src0RC:$src0, clampmod:$clamp),
1597 !if (!eq(NumSrcArgs, 2),
1598 !if (!eq(HasModifiers, 1),
1599 // VOP 2 with modifiers
1600 !if( !eq(HasOMod, 1),
1601 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1602 Src1Mod:$src1_modifiers, Src1RC:$src1,
1603 clampmod:$clamp, omod:$omod),
1604 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1605 Src1Mod:$src1_modifiers, Src1RC:$src1,
1608 // VOP2 without modifiers
1609 !if (!eq(HasIntClamp, 1),
1610 (ins Src0RC:$src0, Src1RC:$src1, clampmod:$clamp),
1611 (ins Src0RC:$src0, Src1RC:$src1))
1614 /* NumSrcArgs == 3 */,
1615 !if (!eq(HasModifiers, 1),
1616 !if (!eq(HasSrc2Mods, 1),
1617 // VOP3 with modifiers
1618 !if (!eq(HasOMod, 1),
1619 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1620 Src1Mod:$src1_modifiers, Src1RC:$src1,
1621 Src2Mod:$src2_modifiers, Src2RC:$src2,
1622 clampmod:$clamp, omod:$omod),
1623 !if (!eq(HasIntClamp, 1),
1624 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1625 Src1Mod:$src1_modifiers, Src1RC:$src1,
1626 Src2Mod:$src2_modifiers, Src2RC:$src2,
1628 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1629 Src1Mod:$src1_modifiers, Src1RC:$src1,
1630 Src2Mod:$src2_modifiers, Src2RC:$src2))),
1631 // VOP3 with modifiers except src2
1632 !if (!eq(HasOMod, 1),
1633 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1634 Src1Mod:$src1_modifiers, Src1RC:$src1,
1635 Src2RC:$src2, clampmod:$clamp, omod:$omod),
1636 !if (!eq(HasIntClamp, 1),
1637 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1638 Src1Mod:$src1_modifiers, Src1RC:$src1,
1639 Src2RC:$src2, clampmod:$clamp),
1640 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1641 Src1Mod:$src1_modifiers, Src1RC:$src1,
1644 // VOP3 without modifiers
1645 !if (!eq(HasIntClamp, 1),
1646 (ins Src0RC:$src0, Src1RC:$src1, Src2RC:$src2, clampmod:$clamp),
1647 (ins Src0RC:$src0, Src1RC:$src1, Src2RC:$src2))
1651 /// XXX - src1 may only allow VGPRs?
1653 // The modifiers (except clamp) are dummy operands for the benefit of
1654 // printing and parsing. They defer their values to looking at the
1655 // srcN_modifiers for what to print.
1656 class getInsVOP3P <RegisterOperand Src0RC, RegisterOperand Src1RC,
1657 RegisterOperand Src2RC, int NumSrcArgs,
1659 Operand Src0Mod, Operand Src1Mod, Operand Src2Mod> {
1660 dag ret = !if (!eq(NumSrcArgs, 2),
1662 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1663 Src1Mod:$src1_modifiers, Src1RC:$src1,
1665 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1666 neg_lo:$neg_lo, neg_hi:$neg_hi),
1667 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1668 Src1Mod:$src1_modifiers, Src1RC:$src1,
1669 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1670 neg_lo:$neg_lo, neg_hi:$neg_hi)),
1671 // else NumSrcArgs == 3
1673 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1674 Src1Mod:$src1_modifiers, Src1RC:$src1,
1675 Src2Mod:$src2_modifiers, Src2RC:$src2,
1677 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1678 neg_lo:$neg_lo, neg_hi:$neg_hi),
1679 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1680 Src1Mod:$src1_modifiers, Src1RC:$src1,
1681 Src2Mod:$src2_modifiers, Src2RC:$src2,
1682 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1683 neg_lo:$neg_lo, neg_hi:$neg_hi))
1687 class getInsVOP3OpSel <RegisterOperand Src0RC,
1688 RegisterOperand Src1RC,
1689 RegisterOperand Src2RC,
1695 dag ret = !if (!eq(NumSrcArgs, 2),
1697 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1698 Src1Mod:$src1_modifiers, Src1RC:$src1,
1701 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1702 Src1Mod:$src1_modifiers, Src1RC:$src1,
1704 // else NumSrcArgs == 3
1706 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1707 Src1Mod:$src1_modifiers, Src1RC:$src1,
1708 Src2Mod:$src2_modifiers, Src2RC:$src2,
1711 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1712 Src1Mod:$src1_modifiers, Src1RC:$src1,
1713 Src2Mod:$src2_modifiers, Src2RC:$src2,
1718 class getInsDPP <RegisterOperand DstRC, RegisterClass Src0RC, RegisterClass Src1RC,
1719 int NumSrcArgs, bit HasModifiers,
1720 Operand Src0Mod, Operand Src1Mod> {
1722 dag ret = !if (!eq(NumSrcArgs, 0),
1723 // VOP1 without input operands (V_NOP)
1724 (ins dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1725 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl),
1726 !if (!eq(NumSrcArgs, 1),
1727 !if (!eq(HasModifiers, 1),
1728 // VOP1_DPP with modifiers
1729 (ins DstRC:$old, Src0Mod:$src0_modifiers,
1730 Src0RC:$src0, dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1731 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl)
1733 // VOP1_DPP without modifiers
1734 (ins DstRC:$old, Src0RC:$src0,
1735 dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1736 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl)
1738 /* NumSrcArgs == 2 */,
1739 !if (!eq(HasModifiers, 1),
1740 // VOP2_DPP with modifiers
1742 Src0Mod:$src0_modifiers, Src0RC:$src0,
1743 Src1Mod:$src1_modifiers, Src1RC:$src1,
1744 dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1745 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl)
1747 // VOP2_DPP without modifiers
1749 Src0RC:$src0, Src1RC:$src1, dpp_ctrl:$dpp_ctrl,
1750 row_mask:$row_mask, bank_mask:$bank_mask,
1751 bound_ctrl:$bound_ctrl)
1755 class getInsDPP16 <RegisterOperand DstRC, RegisterClass Src0RC, RegisterClass Src1RC,
1756 int NumSrcArgs, bit HasModifiers,
1757 Operand Src0Mod, Operand Src1Mod> {
1758 dag ret = !con(getInsDPP<DstRC, Src0RC, Src1RC, NumSrcArgs,
1759 HasModifiers, Src0Mod, Src1Mod>.ret,
1763 class getInsDPP8 <RegisterOperand DstRC, RegisterClass Src0RC, RegisterClass Src1RC,
1764 int NumSrcArgs, bit HasModifiers,
1765 Operand Src0Mod, Operand Src1Mod> {
1766 dag ret = !if (!eq(NumSrcArgs, 0),
1767 // VOP1 without input operands (V_NOP)
1768 (ins dpp8:$dpp8, FI:$fi),
1769 !if (!eq(NumSrcArgs, 1),
1770 !if (!eq(HasModifiers, 1),
1771 // VOP1_DPP with modifiers
1772 (ins DstRC:$old, Src0Mod:$src0_modifiers,
1773 Src0RC:$src0, dpp8:$dpp8, FI:$fi)
1775 // VOP1_DPP without modifiers
1776 (ins DstRC:$old, Src0RC:$src0, dpp8:$dpp8, FI:$fi)
1778 /* NumSrcArgs == 2 */,
1779 !if (!eq(HasModifiers, 1),
1780 // VOP2_DPP with modifiers
1782 Src0Mod:$src0_modifiers, Src0RC:$src0,
1783 Src1Mod:$src1_modifiers, Src1RC:$src1,
1786 // VOP2_DPP without modifiers
1788 Src0RC:$src0, Src1RC:$src1, dpp8:$dpp8, FI:$fi)
1794 class getInsSDWA <RegisterOperand Src0RC, RegisterOperand Src1RC, int NumSrcArgs,
1795 bit HasSDWAOMod, Operand Src0Mod, Operand Src1Mod,
1798 dag ret = !if(!eq(NumSrcArgs, 0),
1799 // VOP1 without input operands (V_NOP)
1801 !if(!eq(NumSrcArgs, 1),
1803 !if(!eq(HasSDWAOMod, 0),
1804 // VOP1_SDWA without omod
1805 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1807 dst_sel:$dst_sel, dst_unused:$dst_unused,
1808 src0_sel:$src0_sel),
1809 // VOP1_SDWA with omod
1810 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1811 clampmod:$clamp, omod:$omod,
1812 dst_sel:$dst_sel, dst_unused:$dst_unused,
1813 src0_sel:$src0_sel)),
1814 !if(!eq(NumSrcArgs, 2),
1815 !if(!eq(DstVT.Size, 1),
1817 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1818 Src1Mod:$src1_modifiers, Src1RC:$src1,
1819 clampmod:$clamp, src0_sel:$src0_sel, src1_sel:$src1_sel),
1821 !if(!eq(HasSDWAOMod, 0),
1822 // VOP2_SDWA without omod
1823 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1824 Src1Mod:$src1_modifiers, Src1RC:$src1,
1826 dst_sel:$dst_sel, dst_unused:$dst_unused,
1827 src0_sel:$src0_sel, src1_sel:$src1_sel),
1828 // VOP2_SDWA with omod
1829 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1830 Src1Mod:$src1_modifiers, Src1RC:$src1,
1831 clampmod:$clamp, omod:$omod,
1832 dst_sel:$dst_sel, dst_unused:$dst_unused,
1833 src0_sel:$src0_sel, src1_sel:$src1_sel))),
1834 (ins)/* endif */)));
1837 // Outs for DPP and SDWA
1838 class getOutsExt <bit HasDst, ValueType DstVT, RegisterOperand DstRCExt> {
1839 dag ret = !if(HasDst,
1840 !if(!eq(DstVT.Size, 1),
1841 (outs), // no dst for VOPC, we use "vcc"-token as dst in SDWA VOPC instructions
1842 (outs DstRCExt:$vdst)),
1847 class getOutsSDWA <bit HasDst, ValueType DstVT, RegisterOperand DstRCSDWA> {
1848 dag ret = !if(HasDst,
1849 !if(!eq(DstVT.Size, 1),
1850 (outs DstRCSDWA:$sdst),
1851 (outs DstRCSDWA:$vdst)),
1855 // Returns the assembly string for the inputs and outputs of a VOP[12C]
1856 // instruction. This does not add the _e32 suffix, so it can be reused
1858 class getAsm32 <bit HasDst, int NumSrcArgs, ValueType DstVT = i32> {
1859 string dst = !if(!eq(DstVT.Size, 1), "$sdst", "$vdst"); // use $sdst for VOPC
1860 string src0 = ", $src0";
1861 string src1 = ", $src1";
1862 string src2 = ", $src2";
1863 string ret = !if(HasDst, dst, "") #
1864 !if(!eq(NumSrcArgs, 1), src0, "") #
1865 !if(!eq(NumSrcArgs, 2), src0#src1, "") #
1866 !if(!eq(NumSrcArgs, 3), src0#src1#src2, "");
1869 // Returns the assembly string for the inputs and outputs of a VOP3
1871 class getAsm64 <bit HasDst, int NumSrcArgs, bit HasIntClamp, bit HasModifiers,
1872 bit HasOMod, ValueType DstVT = i32> {
1873 string dst = !if(!eq(DstVT.Size, 1), "$sdst", "$vdst"); // use $sdst for VOPC
1874 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1875 string src1 = !if(!eq(NumSrcArgs, 1), "",
1876 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1877 " $src1_modifiers,"));
1878 string src2 = !if(!eq(NumSrcArgs, 3), " $src2_modifiers", "");
1879 string iclamp = !if(HasIntClamp, "$clamp", "");
1881 !if(!eq(HasModifiers, 0),
1882 getAsm32<HasDst, NumSrcArgs, DstVT>.ret # iclamp,
1883 dst#", "#src0#src1#src2#"$clamp"#!if(HasOMod, "$omod", ""));
1886 // Returns the assembly string for the inputs and outputs of a VOP3P
1888 class getAsmVOP3P <bit HasDst, int NumSrcArgs, bit HasModifiers,
1889 bit HasClamp, ValueType DstVT = i32> {
1890 string dst = " $vdst";
1891 string src0 = !if(!eq(NumSrcArgs, 1), "$src0", "$src0,");
1892 string src1 = !if(!eq(NumSrcArgs, 1), "",
1893 !if(!eq(NumSrcArgs, 2), " $src1",
1895 string src2 = !if(!eq(NumSrcArgs, 3), " $src2", "");
1897 string mods = !if(HasModifiers, "$neg_lo$neg_hi", "");
1898 string clamp = !if(HasClamp, "$clamp", "");
1900 // Each modifier is printed as an array of bits for each operand, so
1901 // all operands are printed as part of src0_modifiers.
1902 string ret = dst#", "#src0#src1#src2#"$op_sel$op_sel_hi"#mods#clamp;
1905 class getAsmVOP3OpSel <int NumSrcArgs,
1910 string dst = " $vdst";
1912 string isrc0 = !if(!eq(NumSrcArgs, 1), "$src0", "$src0,");
1913 string isrc1 = !if(!eq(NumSrcArgs, 1), "",
1914 !if(!eq(NumSrcArgs, 2), " $src1",
1916 string isrc2 = !if(!eq(NumSrcArgs, 3), " $src2", "");
1918 string fsrc0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1919 string fsrc1 = !if(!eq(NumSrcArgs, 1), "",
1920 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1921 " $src1_modifiers,"));
1922 string fsrc2 = !if(!eq(NumSrcArgs, 3), " $src2_modifiers", "");
1924 string src0 = !if(Src0HasMods, fsrc0, isrc0);
1925 string src1 = !if(Src1HasMods, fsrc1, isrc1);
1926 string src2 = !if(Src2HasMods, fsrc2, isrc2);
1928 string clamp = !if(HasClamp, "$clamp", "");
1930 string ret = dst#", "#src0#src1#src2#"$op_sel"#clamp;
1933 class getAsmDPP <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> {
1934 string dst = !if(HasDst,
1935 !if(!eq(DstVT.Size, 1),
1938 ""); // use $sdst for VOPC
1939 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1940 string src1 = !if(!eq(NumSrcArgs, 1), "",
1941 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1942 " $src1_modifiers,"));
1943 string args = !if(!eq(HasModifiers, 0),
1944 getAsm32<0, NumSrcArgs, DstVT>.ret,
1946 string ret = dst#args#" $dpp_ctrl$row_mask$bank_mask$bound_ctrl";
1949 class getAsmDPP16 <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> {
1950 string ret = getAsmDPP<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret#"$fi";
1953 class getAsmDPP8 <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> {
1954 string dst = !if(HasDst,
1955 !if(!eq(DstVT.Size, 1),
1958 ""); // use $sdst for VOPC
1959 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1960 string src1 = !if(!eq(NumSrcArgs, 1), "",
1961 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1962 " $src1_modifiers,"));
1963 string args = !if(!eq(HasModifiers, 0),
1964 getAsm32<0, NumSrcArgs, DstVT>.ret,
1966 string ret = dst#args#"$dpp8$fi";
1969 class getAsmSDWA <bit HasDst, int NumSrcArgs, ValueType DstVT = i32> {
1970 string dst = !if(HasDst,
1971 !if(!eq(DstVT.Size, 1),
1972 " vcc", // use vcc token as dst for VOPC instructioins
1975 string src0 = "$src0_modifiers";
1976 string src1 = "$src1_modifiers";
1977 string args = !if(!eq(NumSrcArgs, 0),
1979 !if(!eq(NumSrcArgs, 1),
1981 ", "#src0#", "#src1#"$clamp"
1984 string sdwa = !if(!eq(NumSrcArgs, 0),
1986 !if(!eq(NumSrcArgs, 1),
1987 " $dst_sel $dst_unused $src0_sel",
1988 !if(!eq(DstVT.Size, 1),
1989 " $src0_sel $src1_sel", // No dst_sel and dst_unused for VOPC
1990 " $dst_sel $dst_unused $src0_sel $src1_sel"
1994 string ret = dst#args#sdwa;
1997 class getAsmSDWA9 <bit HasDst, bit HasOMod, int NumSrcArgs,
1998 ValueType DstVT = i32> {
1999 string dst = !if(HasDst,
2000 !if(!eq(DstVT.Size, 1),
2004 string src0 = "$src0_modifiers";
2005 string src1 = "$src1_modifiers";
2006 string out_mods = !if(!eq(HasOMod, 0), "$clamp", "$clamp$omod");
2007 string args = !if(!eq(NumSrcArgs, 0), "",
2008 !if(!eq(NumSrcArgs, 1),
2013 string sdwa = !if(!eq(NumSrcArgs, 0), "",
2014 !if(!eq(NumSrcArgs, 1),
2015 out_mods#" $dst_sel $dst_unused $src0_sel",
2016 !if(!eq(DstVT.Size, 1),
2017 " $src0_sel $src1_sel", // No dst_sel, dst_unused and output modifiers for VOPC
2018 out_mods#" $dst_sel $dst_unused $src0_sel $src1_sel"
2022 string ret = dst#args#sdwa;
2026 // Function that checks if instruction supports DPP and SDWA
2027 class getHasExt <int NumSrcArgs, ValueType DstVT = i32, ValueType Src0VT = i32,
2028 ValueType Src1VT = i32> {
2029 bit ret = !if(!eq(NumSrcArgs, 3),
2030 0, // NumSrcArgs == 3 - No DPP or SDWA for VOP3
2031 !if(!eq(DstVT.Size, 64),
2032 0, // 64-bit dst - No DPP or SDWA for 64-bit operands
2033 !if(!eq(Src0VT.Size, 64),
2035 !if(!eq(Src1VT.Size, 64),
2044 class getHasDPP <int NumSrcArgs, ValueType DstVT = i32, ValueType Src0VT = i32,
2045 ValueType Src1VT = i32> {
2046 bit ret = !if(!eq(NumSrcArgs, 0), 0,
2047 getHasExt<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret);
2050 class BitOr<bit a, bit b> {
2051 bit ret = !if(a, 1, !if(b, 1, 0));
2054 class BitAnd<bit a, bit b> {
2055 bit ret = !if(a, !if(b, 1, 0), 0);
2063 class VOPProfile <list<ValueType> _ArgVT, bit _EnableF32SrcMods = 0,
2064 bit _EnableClamp = 0> {
2066 field list<ValueType> ArgVT = _ArgVT;
2067 field bit EnableF32SrcMods = _EnableF32SrcMods;
2068 field bit EnableClamp = _EnableClamp;
2070 field ValueType DstVT = ArgVT[0];
2071 field ValueType Src0VT = ArgVT[1];
2072 field ValueType Src1VT = ArgVT[2];
2073 field ValueType Src2VT = ArgVT[3];
2074 field RegisterOperand DstRC = getVALUDstForVT<DstVT>.ret;
2075 field RegisterOperand DstRCDPP = getVALUDstForVT<DstVT>.ret;
2076 field RegisterOperand DstRCSDWA = getSDWADstForVT<DstVT>.ret;
2077 field RegisterOperand Src0RC32 = getVOPSrc0ForVT<Src0VT>.ret;
2078 field RegisterClass Src1RC32 = getVregSrcForVT<Src1VT>.ret;
2079 field RegisterOperand Src0RC64 = getVOP3SrcForVT<Src0VT>.ret;
2080 field RegisterOperand Src1RC64 = getVOP3SrcForVT<Src1VT>.ret;
2081 field RegisterOperand Src2RC64 = getVOP3SrcForVT<Src2VT>.ret;
2082 field RegisterClass Src0DPP = getVregSrcForVT<Src0VT>.ret;
2083 field RegisterClass Src1DPP = getVregSrcForVT<Src1VT>.ret;
2084 field RegisterOperand Src0SDWA = getSDWASrcForVT<Src0VT>.ret;
2085 field RegisterOperand Src1SDWA = getSDWASrcForVT<Src0VT>.ret;
2086 field Operand Src0Mod = getSrcMod<Src0VT, EnableF32SrcMods>.ret;
2087 field Operand Src1Mod = getSrcMod<Src1VT, EnableF32SrcMods>.ret;
2088 field Operand Src2Mod = getSrcMod<Src2VT, EnableF32SrcMods>.ret;
2089 field Operand Src0ModDPP = getSrcModExt<Src0VT>.ret;
2090 field Operand Src1ModDPP = getSrcModExt<Src1VT>.ret;
2091 field Operand Src0ModSDWA = getSrcModSDWA<Src0VT>.ret;
2092 field Operand Src1ModSDWA = getSrcModSDWA<Src1VT>.ret;
2095 field bit HasDst = !if(!eq(DstVT.Value, untyped.Value), 0, 1);
2096 field bit HasDst32 = HasDst;
2097 field bit EmitDst = HasDst; // force dst encoding, see v_movreld_b32 special case
2098 field int NumSrcArgs = getNumSrcArgs<Src0VT, Src1VT, Src2VT>.ret;
2099 field bit HasSrc0 = !if(!eq(Src0VT.Value, untyped.Value), 0, 1);
2100 field bit HasSrc1 = !if(!eq(Src1VT.Value, untyped.Value), 0, 1);
2101 field bit HasSrc2 = !if(!eq(Src2VT.Value, untyped.Value), 0, 1);
2103 // TODO: Modifiers logic is somewhat adhoc here, to be refined later
2104 // HasModifiers affects the normal and DPP encodings. We take note of EnableF32SrcMods, which
2105 // enables modifiers for i32 type.
2106 field bit HasModifiers = BitOr<isModifierType<Src0VT>.ret, EnableF32SrcMods>.ret;
2108 // HasSrc*FloatMods affects the SDWA encoding. We ignore EnableF32SrcMods.
2109 field bit HasSrc0FloatMods = isFloatType<Src0VT>.ret;
2110 field bit HasSrc1FloatMods = isFloatType<Src1VT>.ret;
2111 field bit HasSrc2FloatMods = isFloatType<Src2VT>.ret;
2113 // HasSrc*IntMods affects the SDWA encoding. We ignore EnableF32SrcMods.
2114 field bit HasSrc0IntMods = isIntType<Src0VT>.ret;
2115 field bit HasSrc1IntMods = isIntType<Src1VT>.ret;
2116 field bit HasSrc2IntMods = isIntType<Src2VT>.ret;
2118 field bit HasSrc0Mods = HasModifiers;
2119 field bit HasSrc1Mods = !if(HasModifiers, BitOr<HasSrc1FloatMods, HasSrc1IntMods>.ret, 0);
2120 field bit HasSrc2Mods = !if(HasModifiers, BitOr<HasSrc2FloatMods, HasSrc2IntMods>.ret, 0);
2122 field bit HasClamp = BitOr<isModifierType<Src0VT>.ret, EnableClamp>.ret;
2123 field bit HasSDWAClamp = EmitDst;
2124 field bit HasFPClamp = BitAnd<isFloatType<DstVT>.ret, HasClamp>.ret;
2125 field bit HasIntClamp = !if(isFloatType<DstVT>.ret, 0, HasClamp);
2126 field bit HasClampLo = HasClamp;
2127 field bit HasClampHi = BitAnd<isPackedType<DstVT>.ret, HasClamp>.ret;
2128 field bit HasHigh = 0;
2130 field bit IsPacked = isPackedType<Src0VT>.ret;
2131 field bit HasOpSel = IsPacked;
2132 field bit HasOMod = !if(HasOpSel, 0, isFloatType<DstVT>.ret);
2133 field bit HasSDWAOMod = isFloatType<DstVT>.ret;
2135 field bit HasExt = getHasExt<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret;
2136 field bit HasExtDPP = getHasDPP<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret;
2137 field bit HasExtSDWA = HasExt;
2138 field bit HasExtSDWA9 = HasExt;
2139 field int NeedPatGen = PatGenMode.NoPattern;
2141 field bit IsMAI = 0;
2142 field bit IsDOT = 0;
2144 field Operand Src0PackedMod = !if(HasSrc0FloatMods, PackedF16InputMods, PackedI16InputMods);
2145 field Operand Src1PackedMod = !if(HasSrc1FloatMods, PackedF16InputMods, PackedI16InputMods);
2146 field Operand Src2PackedMod = !if(HasSrc2FloatMods, PackedF16InputMods, PackedI16InputMods);
2148 field dag Outs = !if(HasDst,(outs DstRC:$vdst),(outs));
2150 // VOP3b instructions are a special case with a second explicit
2151 // output. This is manually overridden for them.
2152 field dag Outs32 = Outs;
2153 field dag Outs64 = Outs;
2154 field dag OutsDPP = getOutsExt<HasDst, DstVT, DstRCDPP>.ret;
2155 field dag OutsDPP8 = getOutsExt<HasDst, DstVT, DstRCDPP>.ret;
2156 field dag OutsSDWA = getOutsSDWA<HasDst, DstVT, DstRCSDWA>.ret;
2158 field dag Ins32 = getIns32<Src0RC32, Src1RC32, NumSrcArgs>.ret;
2159 field dag Ins64 = getIns64<Src0RC64, Src1RC64, Src2RC64, NumSrcArgs,
2160 HasIntClamp, HasModifiers, HasSrc2Mods,
2161 HasOMod, Src0Mod, Src1Mod, Src2Mod>.ret;
2162 field dag InsVOP3P = getInsVOP3P<Src0RC64, Src1RC64, Src2RC64,
2163 NumSrcArgs, HasClamp,
2164 Src0PackedMod, Src1PackedMod, Src2PackedMod>.ret;
2165 field dag InsVOP3OpSel = getInsVOP3OpSel<Src0RC64, Src1RC64, Src2RC64,
2168 getOpSelMod<Src0VT>.ret,
2169 getOpSelMod<Src1VT>.ret,
2170 getOpSelMod<Src2VT>.ret>.ret;
2171 field dag InsDPP = !if(HasExtDPP,
2172 getInsDPP<DstRCDPP, Src0DPP, Src1DPP, NumSrcArgs,
2173 HasModifiers, Src0ModDPP, Src1ModDPP>.ret,
2175 field dag InsDPP16 = getInsDPP16<DstRCDPP, Src0DPP, Src1DPP, NumSrcArgs,
2176 HasModifiers, Src0ModDPP, Src1ModDPP>.ret;
2177 field dag InsDPP8 = getInsDPP8<DstRCDPP, Src0DPP, Src1DPP, NumSrcArgs, 0,
2178 Src0ModDPP, Src1ModDPP>.ret;
2179 field dag InsSDWA = getInsSDWA<Src0SDWA, Src1SDWA, NumSrcArgs,
2180 HasSDWAOMod, Src0ModSDWA, Src1ModSDWA,
2184 field string Asm32 = getAsm32<HasDst, NumSrcArgs, DstVT>.ret;
2185 field string Asm64 = getAsm64<HasDst, NumSrcArgs, HasIntClamp, HasModifiers, HasOMod, DstVT>.ret;
2186 field string AsmVOP3P = getAsmVOP3P<HasDst, NumSrcArgs, HasModifiers, HasClamp, DstVT>.ret;
2187 field string AsmVOP3OpSel = getAsmVOP3OpSel<NumSrcArgs,
2191 HasSrc2FloatMods>.ret;
2192 field string AsmDPP = !if(HasExtDPP,
2193 getAsmDPP<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret, "");
2194 field string AsmDPP16 = getAsmDPP16<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret;
2195 field string AsmDPP8 = getAsmDPP8<HasDst, NumSrcArgs, 0, DstVT>.ret;
2196 field string AsmSDWA = getAsmSDWA<HasDst, NumSrcArgs, DstVT>.ret;
2197 field string AsmSDWA9 = getAsmSDWA9<HasDst, HasSDWAOMod, NumSrcArgs, DstVT>.ret;
2199 field string TieRegDPP = "$old";
2202 class VOP_NO_EXT <VOPProfile p> : VOPProfile <p.ArgVT> {
2206 let HasExtSDWA9 = 0;
2209 class VOP_PAT_GEN <VOPProfile p, int mode=PatGenMode.Pattern> : VOPProfile <p.ArgVT> {
2210 let NeedPatGen = mode;
2213 def VOP_F16_F16 : VOPProfile <[f16, f16, untyped, untyped]>;
2214 def VOP_F16_I16 : VOPProfile <[f16, i16, untyped, untyped]>;
2215 def VOP_I16_F16 : VOPProfile <[i16, f16, untyped, untyped]>;
2217 def VOP_F16_F16_F16 : VOPProfile <[f16, f16, f16, untyped]>;
2218 def VOP_F16_F16_I16 : VOPProfile <[f16, f16, i16, untyped]>;
2219 def VOP_F16_F16_I32 : VOPProfile <[f16, f16, i32, untyped]>;
2220 def VOP_I16_I16_I16 : VOPProfile <[i16, i16, i16, untyped]>;
2222 def VOP_I16_I16_I16_I16 : VOPProfile <[i16, i16, i16, i16, untyped]>;
2223 def VOP_F16_F16_F16_F16 : VOPProfile <[f16, f16, f16, f16, untyped]>;
2225 def VOP_I32_I16_I16_I32 : VOPProfile <[i32, i16, i16, i32, untyped]>;
2227 def VOP_V2F16_V2F16_V2F16 : VOPProfile <[v2f16, v2f16, v2f16, untyped]>;
2228 def VOP_V2I16_V2I16_V2I16 : VOPProfile <[v2i16, v2i16, v2i16, untyped]>;
2229 def VOP_B32_F16_F16 : VOPProfile <[i32, f16, f16, untyped]>;
2231 def VOP_V2F16_V2F16_V2F16_V2F16 : VOPProfile <[v2f16, v2f16, v2f16, v2f16]>;
2232 def VOP_V2I16_V2I16_V2I16_V2I16 : VOPProfile <[v2i16, v2i16, v2i16, v2i16]>;
2233 def VOP_V2I16_F32_F32 : VOPProfile <[v2i16, f32, f32, untyped]>;
2234 def VOP_V2I16_I32_I32 : VOPProfile <[v2i16, i32, i32, untyped]>;
2236 def VOP_F32_V2F16_V2F16_V2F16 : VOPProfile <[f32, v2f16, v2f16, v2f16]>;
2238 def VOP_NONE : VOPProfile <[untyped, untyped, untyped, untyped]>;
2240 def VOP_F32_F32 : VOPProfile <[f32, f32, untyped, untyped]>;
2241 def VOP_F32_F64 : VOPProfile <[f32, f64, untyped, untyped]>;
2242 def VOP_F32_I32 : VOPProfile <[f32, i32, untyped, untyped]>;
2243 def VOP_F64_F32 : VOPProfile <[f64, f32, untyped, untyped]>;
2244 def VOP_F64_F64 : VOPProfile <[f64, f64, untyped, untyped]>;
2245 def VOP_F64_I32 : VOPProfile <[f64, i32, untyped, untyped]>;
2246 def VOP_I32_F32 : VOPProfile <[i32, f32, untyped, untyped]>;
2247 def VOP_I32_F64 : VOPProfile <[i32, f64, untyped, untyped]>;
2248 def VOP_I32_I32 : VOPProfile <[i32, i32, untyped, untyped]>;
2249 def VOP_F16_F32 : VOPProfile <[f16, f32, untyped, untyped]>;
2250 def VOP_F32_F16 : VOPProfile <[f32, f16, untyped, untyped]>;
2252 def VOP_F32_F32_F16 : VOPProfile <[f32, f32, f16, untyped]>;
2253 def VOP_F32_F32_F32 : VOPProfile <[f32, f32, f32, untyped]>;
2254 def VOP_F32_F32_I32 : VOPProfile <[f32, f32, i32, untyped]>;
2255 def VOP_F64_F64_F64 : VOPProfile <[f64, f64, f64, untyped]>;
2256 def VOP_F64_F64_I32 : VOPProfile <[f64, f64, i32, untyped]>;
2257 def VOP_I32_F32_F32 : VOPProfile <[i32, f32, f32, untyped]>;
2258 def VOP_I32_F32_I32 : VOPProfile <[i32, f32, i32, untyped]>;
2259 def VOP_I32_I32_I32 : VOPProfile <[i32, i32, i32, untyped]>;
2260 def VOP_I32_I32_I32_ARITH : VOPProfile <[i32, i32, i32, untyped], 0, /*EnableClamp=*/1>;
2261 def VOP_V2F16_F32_F32 : VOPProfile <[v2f16, f32, f32, untyped]>;
2262 def VOP_F32_F16_F16_F16 : VOPProfile <[f32, f16, f16, f16]>;
2264 def VOP_I64_I64_I32 : VOPProfile <[i64, i64, i32, untyped]>;
2265 def VOP_I64_I32_I64 : VOPProfile <[i64, i32, i64, untyped]>;
2266 def VOP_I64_I64_I64 : VOPProfile <[i64, i64, i64, untyped]>;
2268 def VOP_F16_F32_F16_F32 : VOPProfile <[f16, f32, f16, f32]>;
2269 def VOP_F32_F32_F16_F16 : VOPProfile <[f32, f32, f16, f16]>;
2270 def VOP_F32_F32_F32_F32 : VOPProfile <[f32, f32, f32, f32]>;
2271 def VOP_F64_F64_F64_F64 : VOPProfile <[f64, f64, f64, f64]>;
2272 def VOP_I32_I32_I32_I32 : VOPProfile <[i32, i32, i32, i32]>;
2273 def VOP_I64_I32_I32_I64 : VOPProfile <[i64, i32, i32, i64]>;
2274 def VOP_I32_F32_I32_I32 : VOPProfile <[i32, f32, i32, i32]>;
2275 def VOP_I64_I64_I32_I64 : VOPProfile <[i64, i64, i32, i64]>;
2276 def VOP_V4I32_I64_I32_V4I32 : VOPProfile <[v4i32, i64, i32, v4i32]>;
2278 def VOP_F32_V2F16_V2F16_F32 : VOPProfile <[f32, v2f16, v2f16, f32]>;
2279 def VOP_I32_V2I16_V2I16_I32 : VOPProfile <[i32, v2i16, v2i16, i32]>;
2281 def VOP_V4F32_F32_F32_V4F32 : VOPProfile <[v4f32, f32, f32, v4f32]>;
2282 def VOP_V16F32_F32_F32_V16F32 : VOPProfile <[v16f32, f32, f32, v16f32]>;
2283 def VOP_V32F32_F32_F32_V32F32 : VOPProfile <[v32f32, f32, f32, v32f32]>;
2284 def VOP_V4F32_V4F16_V4F16_V4F32 : VOPProfile <[v4f32, v4f16, v4f16, v4f32]>;
2285 def VOP_V16F32_V4F16_V4F16_V16F32 : VOPProfile <[v16f32, v4f16, v4f16, v16f32]>;
2286 def VOP_V32F32_V4F16_V4F16_V32F32 : VOPProfile <[v32f32, v4f16, v4f16, v32f32]>;
2287 def VOP_V4F32_V2I16_V2I16_V4F32 : VOPProfile <[v4f32, v2i16, v2i16, v4f32]>;
2288 def VOP_V16F32_V2I16_V2I16_V16F32 : VOPProfile <[v16f32, v2i16, v2i16, v16f32]>;
2289 def VOP_V32F32_V2I16_V2I16_V32F32 : VOPProfile <[v32f32, v2i16, v2i16, v32f32]>;
2290 def VOP_V4I32_I32_I32_V4I32 : VOPProfile <[v4i32, i32, i32, v4i32]>;
2291 def VOP_V16I32_I32_I32_V16I32 : VOPProfile <[v16i32, i32, i32, v16i32]>;
2292 def VOP_V32I32_I32_I32_V32I32 : VOPProfile <[v32i32, i32, i32, v32i32]>;
2294 class Commutable_REV <string revOp, bit isOrig> {
2295 string RevOp = revOp;
2296 bit IsOrig = isOrig;
2299 class AtomicNoRet <string noRetOp, bit isRet> {
2300 string NoRetOp = noRetOp;
2304 //===----------------------------------------------------------------------===//
2305 // Interpolation opcodes
2306 //===----------------------------------------------------------------------===//
2308 class VINTRPDstOperand <RegisterClass rc> : RegisterOperand <rc, "printVINTRPDst">;
2310 class VINTRP_Pseudo <string opName, dag outs, dag ins, list<dag> pattern> :
2311 VINTRPCommon <outs, ins, "", pattern>,
2312 SIMCInstr<opName, SIEncodingFamily.NONE> {
2314 let isCodeGenOnly = 1;
2317 // FIXME-GFX10: WIP.
2318 class VINTRP_Real_si <bits <2> op, string opName, dag outs, dag ins,
2319 string asm, int encodingFamily> :
2320 VINTRPCommon <outs, ins, asm, []>,
2322 SIMCInstr<opName, encodingFamily> {
2323 let DisableDecoder = DisableSIDecoder;
2326 class VINTRP_Real_vi <bits <2> op, string opName, dag outs, dag ins,
2328 VINTRPCommon <outs, ins, asm, []>,
2330 SIMCInstr<opName, SIEncodingFamily.VI> {
2331 let AssemblerPredicate = VIAssemblerPredicate;
2332 let DecoderNamespace = "GFX8";
2333 let DisableDecoder = DisableVIDecoder;
2336 // FIXME-GFX10: WIP.
2337 multiclass VINTRP_m <bits <2> op, dag outs, dag ins, string asm,
2338 list<dag> pattern = []> {
2339 def "" : VINTRP_Pseudo <NAME, outs, ins, pattern>;
2341 let AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" in {
2342 def _si : VINTRP_Real_si <op, NAME, outs, ins, asm, SIEncodingFamily.SI>;
2343 } // End AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7"
2345 def _vi : VINTRP_Real_vi <op, NAME, outs, ins, asm>;
2347 let AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" in {
2348 def _gfx10 : VINTRP_Real_si<op, NAME, outs, ins, asm, SIEncodingFamily.GFX10>;
2349 } // End AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10"
2351 //===----------------------------------------------------------------------===//
2352 // Vector instruction mappings
2353 //===----------------------------------------------------------------------===//
2355 // Maps an opcode in e32 form to its e64 equivalent
2356 def getVOPe64 : InstrMapping {
2357 let FilterClass = "VOP";
2358 let RowFields = ["OpName"];
2359 let ColFields = ["Size", "VOP3"];
2360 let KeyCol = ["4", "0"];
2361 let ValueCols = [["8", "1"]];
2364 // Maps an opcode in e64 form to its e32 equivalent
2365 def getVOPe32 : InstrMapping {
2366 let FilterClass = "VOP";
2367 let RowFields = ["OpName"];
2368 let ColFields = ["Size", "VOP3"];
2369 let KeyCol = ["8", "1"];
2370 let ValueCols = [["4", "0"]];
2373 // Maps ordinary instructions to their SDWA counterparts
2374 def getSDWAOp : InstrMapping {
2375 let FilterClass = "VOP";
2376 let RowFields = ["OpName"];
2377 let ColFields = ["AsmVariantName"];
2378 let KeyCol = ["Default"];
2379 let ValueCols = [["SDWA"]];
2382 // Maps SDWA instructions to their ordinary counterparts
2383 def getBasicFromSDWAOp : InstrMapping {
2384 let FilterClass = "VOP";
2385 let RowFields = ["OpName"];
2386 let ColFields = ["AsmVariantName"];
2387 let KeyCol = ["SDWA"];
2388 let ValueCols = [["Default"]];
2391 // Maps ordinary instructions to their DPP counterparts
2392 def getDPPOp32 : InstrMapping {
2393 let FilterClass = "VOP";
2394 let RowFields = ["OpName"];
2395 let ColFields = ["AsmVariantName"];
2396 let KeyCol = ["Default"];
2397 let ValueCols = [["DPP"]];
2400 // Maps an commuted opcode to its original version
2401 def getCommuteOrig : InstrMapping {
2402 let FilterClass = "Commutable_REV";
2403 let RowFields = ["RevOp"];
2404 let ColFields = ["IsOrig"];
2406 let ValueCols = [["1"]];
2409 // Maps an original opcode to its commuted version
2410 def getCommuteRev : InstrMapping {
2411 let FilterClass = "Commutable_REV";
2412 let RowFields = ["RevOp"];
2413 let ColFields = ["IsOrig"];
2415 let ValueCols = [["0"]];
2418 def getMCOpcodeGen : InstrMapping {
2419 let FilterClass = "SIMCInstr";
2420 let RowFields = ["PseudoInstr"];
2421 let ColFields = ["Subtarget"];
2422 let KeyCol = [!cast<string>(SIEncodingFamily.NONE)];
2423 let ValueCols = [[!cast<string>(SIEncodingFamily.SI)],
2424 [!cast<string>(SIEncodingFamily.VI)],
2425 [!cast<string>(SIEncodingFamily.SDWA)],
2426 [!cast<string>(SIEncodingFamily.SDWA9)],
2427 // GFX80 encoding is added to work around a multiple matching
2428 // issue for buffer instructions with unpacked d16 data. This
2429 // does not actually change the encoding, and thus may be
2431 [!cast<string>(SIEncodingFamily.GFX80)],
2432 [!cast<string>(SIEncodingFamily.GFX9)],
2433 [!cast<string>(SIEncodingFamily.GFX10)],
2434 [!cast<string>(SIEncodingFamily.SDWA10)]];
2437 // Get equivalent SOPK instruction.
2438 def getSOPKOp : InstrMapping {
2439 let FilterClass = "SOPKInstTable";
2440 let RowFields = ["BaseCmpOp"];
2441 let ColFields = ["IsSOPK"];
2443 let ValueCols = [["1"]];
2446 def getAddr64Inst : InstrMapping {
2447 let FilterClass = "MUBUFAddr64Table";
2448 let RowFields = ["OpName"];
2449 let ColFields = ["IsAddr64"];
2451 let ValueCols = [["1"]];
2454 def getIfAddr64Inst : InstrMapping {
2455 let FilterClass = "MUBUFAddr64Table";
2456 let RowFields = ["OpName"];
2457 let ColFields = ["IsAddr64"];
2459 let ValueCols = [["1"]];
2462 def getMUBUFNoLdsInst : InstrMapping {
2463 let FilterClass = "MUBUFLdsTable";
2464 let RowFields = ["OpName"];
2465 let ColFields = ["IsLds"];
2467 let ValueCols = [["0"]];
2470 // Maps an atomic opcode to its version with a return value.
2471 def getAtomicRetOp : InstrMapping {
2472 let FilterClass = "AtomicNoRet";
2473 let RowFields = ["NoRetOp"];
2474 let ColFields = ["IsRet"];
2476 let ValueCols = [["1"]];
2479 // Maps an atomic opcode to its returnless version.
2480 def getAtomicNoRetOp : InstrMapping {
2481 let FilterClass = "AtomicNoRet";
2482 let RowFields = ["NoRetOp"];
2483 let ColFields = ["IsRet"];
2485 let ValueCols = [["0"]];
2488 // Maps a GLOBAL to its SADDR form.
2489 def getGlobalSaddrOp : InstrMapping {
2490 let FilterClass = "GlobalSaddrTable";
2491 let RowFields = ["SaddrOp"];
2492 let ColFields = ["IsSaddr"];
2494 let ValueCols = [["1"]];
2497 // Maps a v_cmpx opcode with sdst to opcode without sdst.
2498 def getVCMPXNoSDstOp : InstrMapping {
2499 let FilterClass = "VCMPXNoSDstTable";
2500 let RowFields = ["NoSDstOp"];
2501 let ColFields = ["HasSDst"];
2503 let ValueCols = [["0"]];
2506 // Maps a SOPP to a SOPP with S_NOP
2507 def getSOPPWithRelaxation : InstrMapping {
2508 let FilterClass = "Base_SOPP";
2509 let RowFields = ["AsmString"];
2510 let ColFields = ["Size"];
2512 let ValueCols = [["8"]];
2515 include "SIInstructions.td"
2517 include "DSInstructions.td"
2518 include "MIMGInstructions.td"