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_i32timm: SDNodeXForm<timm, [{
693 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32);
696 def as_i64imm: SDNodeXForm<imm, [{
697 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i64);
700 def cond_as_i32imm: SDNodeXForm<cond, [{
701 return CurDAG->getTargetConstant(N->get(), SDLoc(N), MVT::i32);
704 // Copied from the AArch64 backend:
705 def bitcast_fpimm_to_i32 : SDNodeXForm<fpimm, [{
706 return CurDAG->getTargetConstant(
707 N->getValueAPF().bitcastToAPInt().getZExtValue(), SDLoc(N), MVT::i32);
710 def frameindex_to_targetframeindex : SDNodeXForm<frameindex, [{
711 auto FI = cast<FrameIndexSDNode>(N);
712 return CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
715 // Copied from the AArch64 backend:
716 def bitcast_fpimm_to_i64 : SDNodeXForm<fpimm, [{
717 return CurDAG->getTargetConstant(
718 N->getValueAPF().bitcastToAPInt().getZExtValue(), SDLoc(N), MVT::i64);
721 class bitextract_imm<int bitnum> : SDNodeXForm<imm, [{
722 uint64_t Imm = N->getZExtValue();
723 unsigned Bit = (Imm >> }] # bitnum # [{ ) & 1;
724 return CurDAG->getTargetConstant(Bit, SDLoc(N), MVT::i1);
727 def SIMM16bit : ImmLeaf <i32,
728 [{return isInt<16>(Imm);}]
731 def UIMM16bit : ImmLeaf <i32,
732 [{return isUInt<16>(Imm);}]
735 def i64imm_32bit : ImmLeaf<i64, [{
736 return (Imm & 0xffffffffULL) == static_cast<uint64_t>(Imm);
739 class InlineImm <ValueType vt> : PatLeaf <(vt imm), [{
740 return isInlineImmediate(N);
743 class InlineFPImm <ValueType vt> : PatLeaf <(vt fpimm), [{
744 return isInlineImmediate(N);
747 class VGPRImm <dag frag> : PatLeaf<frag, [{
751 def NegateImm : SDNodeXForm<imm, [{
752 return CurDAG->getConstant(-N->getSExtValue(), SDLoc(N), MVT::i32);
755 // TODO: When FP inline imm values work?
756 def NegSubInlineConst32 : ImmLeaf<i32, [{
757 return Imm < -16 && Imm >= -64;
760 def NegSubInlineConst16 : ImmLeaf<i16, [{
761 return Imm < -16 && Imm >= -64;
764 def ShiftAmt32Imm : PatLeaf <(imm), [{
765 return N->getZExtValue() < 32;
768 def getNegV2I16Imm : SDNodeXForm<build_vector, [{
769 return SDValue(packNegConstantV2I16(N, *CurDAG), 0);
772 def NegSubInlineConstV216 : PatLeaf<(build_vector), [{
773 assert(N->getNumOperands() == 2);
774 assert(N->getOperand(0).getValueType().getSizeInBits() == 16);
775 SDValue Src0 = N->getOperand(0);
776 SDValue Src1 = N->getOperand(1);
778 return isNegInlineImmediate(Src0.getNode());
780 return (isNullConstantOrUndef(Src0) && isNegInlineImmediate(Src1.getNode())) ||
781 (isNullConstantOrUndef(Src1) && isNegInlineImmediate(Src0.getNode()));
784 //===----------------------------------------------------------------------===//
786 //===----------------------------------------------------------------------===//
788 def SoppBrTarget : AsmOperandClass {
789 let Name = "SoppBrTarget";
790 let ParserMethod = "parseSOppBrTarget";
793 def sopp_brtarget : Operand<OtherVT> {
794 let EncoderMethod = "getSOPPBrEncoding";
795 let DecoderMethod = "decodeSoppBrTarget";
796 let OperandType = "OPERAND_PCREL";
797 let ParserMatchClass = SoppBrTarget;
800 def si_ga : Operand<iPTR>;
802 def InterpSlotMatchClass : AsmOperandClass {
803 let Name = "InterpSlot";
804 let PredicateMethod = "isInterpSlot";
805 let ParserMethod = "parseInterpSlot";
806 let RenderMethod = "addImmOperands";
809 def InterpSlot : Operand<i32> {
810 let PrintMethod = "printInterpSlot";
811 let ParserMatchClass = InterpSlotMatchClass;
812 let OperandType = "OPERAND_IMMEDIATE";
815 def AttrMatchClass : AsmOperandClass {
817 let PredicateMethod = "isInterpAttr";
818 let ParserMethod = "parseInterpAttr";
819 let RenderMethod = "addImmOperands";
822 // It appears to be necessary to create a separate operand for this to
823 // be able to parse attr<num> with no space.
824 def Attr : Operand<i32> {
825 let PrintMethod = "printInterpAttr";
826 let ParserMatchClass = AttrMatchClass;
827 let OperandType = "OPERAND_IMMEDIATE";
830 def AttrChanMatchClass : AsmOperandClass {
831 let Name = "AttrChan";
832 let PredicateMethod = "isAttrChan";
833 let RenderMethod = "addImmOperands";
836 def AttrChan : Operand<i32> {
837 let PrintMethod = "printInterpAttrChan";
838 let ParserMatchClass = AttrChanMatchClass;
839 let OperandType = "OPERAND_IMMEDIATE";
842 def SendMsgMatchClass : AsmOperandClass {
843 let Name = "SendMsg";
844 let PredicateMethod = "isSendMsg";
845 let ParserMethod = "parseSendMsgOp";
846 let RenderMethod = "addImmOperands";
849 def SwizzleMatchClass : AsmOperandClass {
850 let Name = "Swizzle";
851 let PredicateMethod = "isSwizzle";
852 let ParserMethod = "parseSwizzleOp";
853 let RenderMethod = "addImmOperands";
857 def EndpgmMatchClass : AsmOperandClass {
858 let Name = "EndpgmImm";
859 let PredicateMethod = "isEndpgm";
860 let ParserMethod = "parseEndpgmOp";
861 let RenderMethod = "addImmOperands";
865 def ExpTgtMatchClass : AsmOperandClass {
867 let PredicateMethod = "isExpTgt";
868 let ParserMethod = "parseExpTgt";
869 let RenderMethod = "printExpTgt";
872 def SendMsgImm : Operand<i32> {
873 let PrintMethod = "printSendMsg";
874 let ParserMatchClass = SendMsgMatchClass;
877 def SwizzleImm : Operand<i16> {
878 let PrintMethod = "printSwizzle";
879 let ParserMatchClass = SwizzleMatchClass;
882 def EndpgmImm : Operand<i16> {
883 let PrintMethod = "printEndpgm";
884 let ParserMatchClass = EndpgmMatchClass;
887 def SWaitMatchClass : AsmOperandClass {
888 let Name = "SWaitCnt";
889 let RenderMethod = "addImmOperands";
890 let ParserMethod = "parseSWaitCntOps";
893 def VReg32OrOffClass : AsmOperandClass {
894 let Name = "VReg32OrOff";
895 let ParserMethod = "parseVReg32OrOff";
898 def WAIT_FLAG : Operand <i32> {
899 let ParserMatchClass = SWaitMatchClass;
900 let PrintMethod = "printWaitFlag";
901 let OperandType = "OPERAND_IMMEDIATE";
904 include "SIInstrFormats.td"
905 include "VIInstrFormats.td"
907 def BoolReg : AsmOperandClass {
908 let Name = "BoolReg";
909 let ParserMethod = "parseBoolReg";
910 let RenderMethod = "addRegOperands";
913 class BoolRC : RegisterOperand<SReg_1> {
914 let ParserMatchClass = BoolReg;
915 let DecoderMethod = "decodeBoolReg";
918 def SSrc_i1 : RegisterOperand<SReg_1_XEXEC> {
919 let ParserMatchClass = BoolReg;
920 let DecoderMethod = "decodeBoolReg";
923 def VOPDstS64orS32 : BoolRC {
924 let PrintMethod = "printVOPDst";
927 // SCSrc_i1 is the operand for pseudo instructions only.
928 // Boolean immeadiates shall not be exposed to codegen instructions.
929 def SCSrc_i1 : RegisterOperand<SReg_1_XEXEC> {
930 let OperandNamespace = "AMDGPU";
931 let OperandType = "OPERAND_REG_IMM_INT32";
932 let ParserMatchClass = BoolReg;
933 let DecoderMethod = "decodeBoolReg";
936 // ===----------------------------------------------------------------------===//
937 // ExpSrc* Special cases for exp src operands which are printed as
938 // "off" depending on en operand.
939 // ===----------------------------------------------------------------------===//
941 def ExpSrc0 : RegisterOperand<VGPR_32> {
942 let PrintMethod = "printExpSrc0";
943 let ParserMatchClass = VReg32OrOffClass;
946 def ExpSrc1 : RegisterOperand<VGPR_32> {
947 let PrintMethod = "printExpSrc1";
948 let ParserMatchClass = VReg32OrOffClass;
951 def ExpSrc2 : RegisterOperand<VGPR_32> {
952 let PrintMethod = "printExpSrc2";
953 let ParserMatchClass = VReg32OrOffClass;
956 def ExpSrc3 : RegisterOperand<VGPR_32> {
957 let PrintMethod = "printExpSrc3";
958 let ParserMatchClass = VReg32OrOffClass;
961 class SDWASrc<ValueType vt> : RegisterOperand<VS_32> {
962 let OperandNamespace = "AMDGPU";
963 string Type = !if(isFloatType<vt>.ret, "FP", "INT");
964 let OperandType = "OPERAND_REG_INLINE_C_"#Type#vt.Size;
965 let DecoderMethod = "decodeSDWASrc"#vt.Size;
966 let EncoderMethod = "getSDWASrcEncoding";
969 def SDWASrc_i32 : SDWASrc<i32>;
970 def SDWASrc_i16 : SDWASrc<i16>;
971 def SDWASrc_f32 : SDWASrc<f32>;
972 def SDWASrc_f16 : SDWASrc<f16>;
974 def SDWAVopcDst : BoolRC {
975 let OperandNamespace = "AMDGPU";
976 let OperandType = "OPERAND_SDWA_VOPC_DST";
977 let EncoderMethod = "getSDWAVopcDstEncoding";
978 let DecoderMethod = "decodeSDWAVopcDst";
979 let PrintMethod = "printVOPDst";
982 class NamedMatchClass<string CName, bit Optional = 1> : AsmOperandClass {
983 let Name = "Imm"#CName;
984 let PredicateMethod = "is"#CName;
985 let ParserMethod = !if(Optional, "parseOptionalOperand", "parse"#CName);
986 let RenderMethod = "addImmOperands";
987 let IsOptional = Optional;
988 let DefaultMethod = !if(Optional, "default"#CName, ?);
991 class NamedOperandBit<string Name, AsmOperandClass MatchClass> : Operand<i1> {
992 let PrintMethod = "print"#Name;
993 let ParserMatchClass = MatchClass;
996 class NamedOperandU8<string Name, AsmOperandClass MatchClass> : Operand<i8> {
997 let PrintMethod = "print"#Name;
998 let ParserMatchClass = MatchClass;
1001 class NamedOperandU16<string Name, AsmOperandClass MatchClass> : Operand<i16> {
1002 let PrintMethod = "print"#Name;
1003 let ParserMatchClass = MatchClass;
1006 class NamedOperandU32<string Name, AsmOperandClass MatchClass> : Operand<i32> {
1007 let PrintMethod = "print"#Name;
1008 let ParserMatchClass = MatchClass;
1011 class NamedOperandU32Default0<string Name, AsmOperandClass MatchClass> :
1012 OperandWithDefaultOps<i32, (ops (i32 0))> {
1013 let PrintMethod = "print"#Name;
1014 let ParserMatchClass = MatchClass;
1017 let OperandType = "OPERAND_IMMEDIATE" in {
1019 def offen : NamedOperandBit<"Offen", NamedMatchClass<"Offen">>;
1020 def idxen : NamedOperandBit<"Idxen", NamedMatchClass<"Idxen">>;
1021 def addr64 : NamedOperandBit<"Addr64", NamedMatchClass<"Addr64">>;
1023 def flat_offset : NamedOperandU16<"FlatOffset", NamedMatchClass<"FlatOffset">>;
1024 def offset : NamedOperandU16<"Offset", NamedMatchClass<"Offset">>;
1025 def offset0 : NamedOperandU8<"Offset0", NamedMatchClass<"Offset0">>;
1026 def offset1 : NamedOperandU8<"Offset1", NamedMatchClass<"Offset1">>;
1028 def gds : NamedOperandBit<"GDS", NamedMatchClass<"GDS">>;
1030 def omod : NamedOperandU32<"OModSI", NamedMatchClass<"OModSI">>;
1031 def clampmod : NamedOperandBit<"ClampSI", NamedMatchClass<"ClampSI">>;
1032 def highmod : NamedOperandBit<"High", NamedMatchClass<"High">>;
1034 def DLC : NamedOperandBit<"DLC", NamedMatchClass<"DLC">>;
1035 def GLC : NamedOperandBit<"GLC", NamedMatchClass<"GLC">>;
1036 def SLC : NamedOperandBit<"SLC", NamedMatchClass<"SLC">>;
1037 def TFE : NamedOperandBit<"TFE", NamedMatchClass<"TFE">>;
1038 def UNorm : NamedOperandBit<"UNorm", NamedMatchClass<"UNorm">>;
1039 def DA : NamedOperandBit<"DA", NamedMatchClass<"DA">>;
1040 def R128A16 : NamedOperandBit<"R128A16", NamedMatchClass<"R128A16">>;
1041 def D16 : NamedOperandBit<"D16", NamedMatchClass<"D16">>;
1042 def LWE : NamedOperandBit<"LWE", NamedMatchClass<"LWE">>;
1043 def exp_compr : NamedOperandBit<"ExpCompr", NamedMatchClass<"ExpCompr">>;
1044 def exp_vm : NamedOperandBit<"ExpVM", NamedMatchClass<"ExpVM">>;
1046 def FORMAT : NamedOperandU8<"FORMAT", NamedMatchClass<"FORMAT">>;
1048 def DMask : NamedOperandU16<"DMask", NamedMatchClass<"DMask">>;
1049 def Dim : NamedOperandU8<"Dim", NamedMatchClass<"Dim", 0>>;
1051 def dpp8 : NamedOperandU32<"DPP8", NamedMatchClass<"DPP8", 0>>;
1053 def dpp_ctrl : NamedOperandU32<"DPPCtrl", NamedMatchClass<"DPPCtrl", 0>>;
1054 def row_mask : NamedOperandU32<"RowMask", NamedMatchClass<"RowMask">>;
1055 def bank_mask : NamedOperandU32<"BankMask", NamedMatchClass<"BankMask">>;
1056 def bound_ctrl : NamedOperandBit<"BoundCtrl", NamedMatchClass<"BoundCtrl">>;
1057 def FI : NamedOperandU32<"FI", NamedMatchClass<"FI">>;
1059 def dst_sel : NamedOperandU32<"SDWADstSel", NamedMatchClass<"SDWADstSel">>;
1060 def src0_sel : NamedOperandU32<"SDWASrc0Sel", NamedMatchClass<"SDWASrc0Sel">>;
1061 def src1_sel : NamedOperandU32<"SDWASrc1Sel", NamedMatchClass<"SDWASrc1Sel">>;
1062 def dst_unused : NamedOperandU32<"SDWADstUnused", NamedMatchClass<"SDWADstUnused">>;
1064 def op_sel : NamedOperandU32Default0<"OpSel", NamedMatchClass<"OpSel">>;
1065 def op_sel_hi : NamedOperandU32Default0<"OpSelHi", NamedMatchClass<"OpSelHi">>;
1066 def neg_lo : NamedOperandU32Default0<"NegLo", NamedMatchClass<"NegLo">>;
1067 def neg_hi : NamedOperandU32Default0<"NegHi", NamedMatchClass<"NegHi">>;
1069 def blgp : NamedOperandU32<"BLGP", NamedMatchClass<"BLGP">>;
1070 def cbsz : NamedOperandU32<"CBSZ", NamedMatchClass<"CBSZ">>;
1071 def abid : NamedOperandU32<"ABID", NamedMatchClass<"ABID">>;
1073 def hwreg : NamedOperandU16<"Hwreg", NamedMatchClass<"Hwreg", 0>>;
1075 def exp_tgt : NamedOperandU8<"ExpTgt", NamedMatchClass<"ExpTgt", 0>> {
1079 } // End OperandType = "OPERAND_IMMEDIATE"
1081 class KImmMatchClass<int size> : AsmOperandClass {
1082 let Name = "KImmFP"#size;
1083 let PredicateMethod = "isKImmFP"#size;
1084 let ParserMethod = "parseImm";
1085 let RenderMethod = "addKImmFP"#size#"Operands";
1088 class kimmOperand<ValueType vt> : Operand<vt> {
1089 let OperandNamespace = "AMDGPU";
1090 let OperandType = "OPERAND_KIMM"#vt.Size;
1091 let PrintMethod = "printU"#vt.Size#"ImmOperand";
1092 let ParserMatchClass = !cast<AsmOperandClass>("KImmFP"#vt.Size#"MatchClass");
1095 // 32-bit VALU immediate operand that uses the constant bus.
1096 def KImmFP32MatchClass : KImmMatchClass<32>;
1097 def f32kimm : kimmOperand<i32>;
1099 // 32-bit VALU immediate operand with a 16-bit value that uses the
1101 def KImmFP16MatchClass : KImmMatchClass<16>;
1102 def f16kimm : kimmOperand<i16>;
1104 class FPInputModsMatchClass <int opSize> : AsmOperandClass {
1105 let Name = "RegOrImmWithFP"#opSize#"InputMods";
1106 let ParserMethod = "parseRegOrImmWithFPInputMods";
1107 let PredicateMethod = "isRegOrImmWithFP"#opSize#"InputMods";
1110 def FP16InputModsMatchClass : FPInputModsMatchClass<16>;
1111 def FP32InputModsMatchClass : FPInputModsMatchClass<32>;
1112 def FP64InputModsMatchClass : FPInputModsMatchClass<64>;
1114 class InputMods <AsmOperandClass matchClass> : Operand <i32> {
1115 let OperandNamespace = "AMDGPU";
1116 let OperandType = "OPERAND_INPUT_MODS";
1117 let ParserMatchClass = matchClass;
1120 class FPInputMods <FPInputModsMatchClass matchClass> : InputMods <matchClass> {
1121 let PrintMethod = "printOperandAndFPInputMods";
1124 def FP16InputMods : FPInputMods<FP16InputModsMatchClass>;
1125 def FP32InputMods : FPInputMods<FP32InputModsMatchClass>;
1126 def FP64InputMods : FPInputMods<FP64InputModsMatchClass>;
1128 class IntInputModsMatchClass <int opSize> : AsmOperandClass {
1129 let Name = "RegOrImmWithInt"#opSize#"InputMods";
1130 let ParserMethod = "parseRegOrImmWithIntInputMods";
1131 let PredicateMethod = "isRegOrImmWithInt"#opSize#"InputMods";
1133 def Int32InputModsMatchClass : IntInputModsMatchClass<32>;
1134 def Int64InputModsMatchClass : IntInputModsMatchClass<64>;
1136 class IntInputMods <IntInputModsMatchClass matchClass> : InputMods <matchClass> {
1137 let PrintMethod = "printOperandAndIntInputMods";
1139 def Int32InputMods : IntInputMods<Int32InputModsMatchClass>;
1140 def Int64InputMods : IntInputMods<Int64InputModsMatchClass>;
1142 class OpSelModsMatchClass : AsmOperandClass {
1143 let Name = "OpSelMods";
1144 let ParserMethod = "parseRegOrImm";
1145 let PredicateMethod = "isRegOrImm";
1148 def IntOpSelModsMatchClass : OpSelModsMatchClass;
1149 def IntOpSelMods : InputMods<IntOpSelModsMatchClass>;
1151 class FPSDWAInputModsMatchClass <int opSize> : AsmOperandClass {
1152 let Name = "SDWAWithFP"#opSize#"InputMods";
1153 let ParserMethod = "parseRegOrImmWithFPInputMods";
1154 let PredicateMethod = "isSDWAFP"#opSize#"Operand";
1157 def FP16SDWAInputModsMatchClass : FPSDWAInputModsMatchClass<16>;
1158 def FP32SDWAInputModsMatchClass : FPSDWAInputModsMatchClass<32>;
1160 class FPSDWAInputMods <FPSDWAInputModsMatchClass matchClass> :
1161 InputMods <matchClass> {
1162 let PrintMethod = "printOperandAndFPInputMods";
1165 def FP16SDWAInputMods : FPSDWAInputMods<FP16SDWAInputModsMatchClass>;
1166 def FP32SDWAInputMods : FPSDWAInputMods<FP32SDWAInputModsMatchClass>;
1168 def FPVRegInputModsMatchClass : AsmOperandClass {
1169 let Name = "VRegWithFPInputMods";
1170 let ParserMethod = "parseRegWithFPInputMods";
1171 let PredicateMethod = "isVReg32";
1174 def FPVRegInputMods : InputMods <FPVRegInputModsMatchClass> {
1175 let PrintMethod = "printOperandAndFPInputMods";
1178 class IntSDWAInputModsMatchClass <int opSize> : AsmOperandClass {
1179 let Name = "SDWAWithInt"#opSize#"InputMods";
1180 let ParserMethod = "parseRegOrImmWithIntInputMods";
1181 let PredicateMethod = "isSDWAInt"#opSize#"Operand";
1184 def Int16SDWAInputModsMatchClass : IntSDWAInputModsMatchClass<16>;
1185 def Int32SDWAInputModsMatchClass : IntSDWAInputModsMatchClass<32>;
1187 class IntSDWAInputMods <IntSDWAInputModsMatchClass matchClass> :
1188 InputMods <matchClass> {
1189 let PrintMethod = "printOperandAndIntInputMods";
1192 def Int16SDWAInputMods : IntSDWAInputMods<Int16SDWAInputModsMatchClass>;
1193 def Int32SDWAInputMods : IntSDWAInputMods<Int32SDWAInputModsMatchClass>;
1195 def IntVRegInputModsMatchClass : AsmOperandClass {
1196 let Name = "VRegWithIntInputMods";
1197 let ParserMethod = "parseRegWithIntInputMods";
1198 let PredicateMethod = "isVReg32";
1201 def IntVRegInputMods : InputMods <IntVRegInputModsMatchClass> {
1202 let PrintMethod = "printOperandAndIntInputMods";
1205 class PackedFPInputModsMatchClass <int opSize> : AsmOperandClass {
1206 let Name = "PackedFP"#opSize#"InputMods";
1207 let ParserMethod = "parseRegOrImm";
1208 let PredicateMethod = "isRegOrImm";
1209 // let PredicateMethod = "isPackedFP"#opSize#"InputMods";
1212 class PackedIntInputModsMatchClass <int opSize> : AsmOperandClass {
1213 let Name = "PackedInt"#opSize#"InputMods";
1214 let ParserMethod = "parseRegOrImm";
1215 let PredicateMethod = "isRegOrImm";
1216 // let PredicateMethod = "isPackedInt"#opSize#"InputMods";
1219 def PackedF16InputModsMatchClass : PackedFPInputModsMatchClass<16>;
1220 def PackedI16InputModsMatchClass : PackedIntInputModsMatchClass<16>;
1222 class PackedFPInputMods <PackedFPInputModsMatchClass matchClass> : InputMods <matchClass> {
1223 // let PrintMethod = "printPackedFPInputMods";
1226 class PackedIntInputMods <PackedIntInputModsMatchClass matchClass> : InputMods <matchClass> {
1227 //let PrintMethod = "printPackedIntInputMods";
1230 def PackedF16InputMods : PackedFPInputMods<PackedF16InputModsMatchClass>;
1231 def PackedI16InputMods : PackedIntInputMods<PackedI16InputModsMatchClass>;
1233 //===----------------------------------------------------------------------===//
1235 //===----------------------------------------------------------------------===//
1237 def DS1Addr1Offset : ComplexPattern<i32, 2, "SelectDS1Addr1Offset">;
1238 def DS64Bit4ByteAligned : ComplexPattern<i32, 3, "SelectDS64Bit4ByteAligned">;
1240 def MOVRELOffset : ComplexPattern<i32, 2, "SelectMOVRELOffset">;
1242 def VOP3Mods0 : ComplexPattern<untyped, 4, "SelectVOP3Mods0">;
1243 def VOP3Mods0Clamp : ComplexPattern<untyped, 3, "SelectVOP3Mods0Clamp">;
1244 def VOP3Mods0Clamp0OMod : ComplexPattern<untyped, 4, "SelectVOP3Mods0Clamp0OMod">;
1245 def VOP3Mods : ComplexPattern<untyped, 2, "SelectVOP3Mods">;
1246 def VOP3NoMods : ComplexPattern<untyped, 1, "SelectVOP3NoMods">;
1247 // VOP3Mods, but the input source is known to never be NaN.
1248 def VOP3Mods_nnan : ComplexPattern<fAny, 2, "SelectVOP3Mods_NNaN">;
1249 // VOP3Mods, but only allowed for f32 operands.
1250 def VOP3Mods_f32 : ComplexPattern<fAny, 2, "SelectVOP3Mods_f32">;
1252 def VOP3OMods : ComplexPattern<untyped, 3, "SelectVOP3OMods">;
1254 def VOP3PMods : ComplexPattern<untyped, 2, "SelectVOP3PMods">;
1255 def VOP3PMods0 : ComplexPattern<untyped, 3, "SelectVOP3PMods0">;
1257 def VOP3OpSel : ComplexPattern<untyped, 2, "SelectVOP3OpSel">;
1258 def VOP3OpSel0 : ComplexPattern<untyped, 3, "SelectVOP3OpSel0">;
1260 def VOP3OpSelMods : ComplexPattern<untyped, 2, "SelectVOP3OpSelMods">;
1261 def VOP3OpSelMods0 : ComplexPattern<untyped, 3, "SelectVOP3OpSelMods0">;
1263 def VOP3PMadMixMods : ComplexPattern<untyped, 2, "SelectVOP3PMadMixMods">;
1266 def Hi16Elt : ComplexPattern<untyped, 1, "SelectHi16Elt">;
1268 //===----------------------------------------------------------------------===//
1269 // SI assembler operands
1270 //===----------------------------------------------------------------------===//
1275 int FLAT_SCR = 0x68;
1278 // This should be kept in sync with SISrcMods enum
1302 int LLVM_DEBUG_TRAP = 3;
1318 int FLAT_SCR_LO = 20;
1319 int FLAT_SCR_HI = 21;
1320 int XNACK_MASK = 22;
1321 int POPS_PACKER = 25;
1324 class getHwRegImm<int Reg, int Offset = 0, int Size = 32> {
1326 !or(!shl(Offset, 6),
1327 !shl(!add(Size, -1), 11)));
1330 //===----------------------------------------------------------------------===//
1332 // SI Instruction multiclass helpers.
1334 // Instructions with _32 take 32-bit operands.
1335 // Instructions with _64 take 64-bit operands.
1337 // VOP_* instructions can use either a 32-bit or 64-bit encoding. The 32-bit
1338 // encoding is the standard encoding, but instruction that make use of
1339 // any of the instruction modifiers must use the 64-bit encoding.
1341 // Instructions with _e32 use the 32-bit encoding.
1342 // Instructions with _e64 use the 64-bit encoding.
1344 //===----------------------------------------------------------------------===//
1346 class SIMCInstr <string pseudo, int subtarget> {
1347 string PseudoInstr = pseudo;
1348 int Subtarget = subtarget;
1351 //===----------------------------------------------------------------------===//
1353 //===----------------------------------------------------------------------===//
1355 class EXP_Helper<bit done, SDPatternOperator node = null_frag> : EXPCommon<
1358 ExpSrc0:$src0, ExpSrc1:$src1, ExpSrc2:$src2, ExpSrc3:$src3,
1359 exp_vm:$vm, exp_compr:$compr, i8imm:$en),
1360 "exp$tgt $src0, $src1, $src2, $src3"#!if(done, " done", "")#"$compr$vm",
1361 [(node (i8 timm:$tgt), (i8 timm:$en),
1362 f32:$src0, f32:$src1, f32:$src2, f32:$src3,
1363 (i1 timm:$compr), (i1 timm:$vm))]> {
1364 let AsmMatchConverter = "cvtExp";
1367 // Split EXP instruction into EXP and EXP_DONE so we can set
1368 // mayLoad for done=1.
1369 multiclass EXP_m<bit done, SDPatternOperator node> {
1370 let mayLoad = done, DisableWQM = 1 in {
1371 let isPseudo = 1, isCodeGenOnly = 1 in {
1372 def "" : EXP_Helper<done, node>,
1373 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.NONE>;
1376 let done = done in {
1377 def _si : EXP_Helper<done>,
1378 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.SI>,
1380 let AssemblerPredicates = [isGFX6GFX7];
1381 let DecoderNamespace = "GFX6GFX7";
1382 let DisableDecoder = DisableSIDecoder;
1385 def _vi : EXP_Helper<done>,
1386 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.VI>,
1388 let AssemblerPredicates = [isGFX8GFX9];
1389 let DecoderNamespace = "GFX8";
1390 let DisableDecoder = DisableVIDecoder;
1393 def _gfx10 : EXP_Helper<done>,
1394 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.GFX10>,
1396 let AssemblerPredicates = [isGFX10Plus];
1397 let DecoderNamespace = "GFX10";
1398 let DisableDecoder = DisableSIDecoder;
1404 //===----------------------------------------------------------------------===//
1405 // Vector ALU classes
1406 //===----------------------------------------------------------------------===//
1408 class getNumSrcArgs<ValueType Src0, ValueType Src1, ValueType Src2> {
1410 !if (!eq(Src0.Value, untyped.Value), 0,
1411 !if (!eq(Src1.Value, untyped.Value), 1, // VOP1
1412 !if (!eq(Src2.Value, untyped.Value), 2, // VOP2
1416 // Returns the register class to use for the destination of VOP[123C]
1417 // instructions for the given VT.
1418 class getVALUDstForVT<ValueType VT> {
1419 RegisterOperand ret = !if(!eq(VT.Size, 32), VOPDstOperand<VGPR_32>,
1420 !if(!eq(VT.Size, 128), VOPDstOperand<VReg_128>,
1421 !if(!eq(VT.Size, 64), VOPDstOperand<VReg_64>,
1422 !if(!eq(VT.Size, 16), VOPDstOperand<VGPR_32>,
1423 VOPDstS64orS32)))); // else VT == i1
1426 // Returns true if VT is floating point.
1427 class getIsFP<ValueType VT> {
1428 bit ret = !if(!eq(VT.Value, f16.Value), 1,
1429 !if(!eq(VT.Value, v2f16.Value), 1,
1430 !if(!eq(VT.Value, v4f16.Value), 1,
1431 !if(!eq(VT.Value, f32.Value), 1,
1432 !if(!eq(VT.Value, v2f32.Value), 1,
1433 !if(!eq(VT.Value, f64.Value), 1,
1434 !if(!eq(VT.Value, v2f64.Value), 1,
1438 // Returns the register class to use for the destination of VOP[12C]
1439 // instructions with SDWA extension
1440 class getSDWADstForVT<ValueType VT> {
1441 RegisterOperand ret = !if(!eq(VT.Size, 1),
1442 SDWAVopcDst, // VOPC
1443 VOPDstOperand<VGPR_32>); // VOP1/2 32-bit dst
1446 // Returns the register class to use for source 0 of VOP[12C]
1447 // instructions for the given VT.
1448 class getVOPSrc0ForVT<ValueType VT> {
1449 bit isFP = getIsFP<VT>.ret;
1451 RegisterOperand ret =
1453 !if(!eq(VT.Size, 64),
1455 !if(!eq(VT.Value, f16.Value),
1457 !if(!eq(VT.Value, v2f16.Value),
1459 !if(!eq(VT.Value, v4f16.Value),
1466 !if(!eq(VT.Size, 64),
1468 !if(!eq(VT.Value, i16.Value),
1470 !if(!eq(VT.Value, v2i16.Value),
1479 // Returns the vreg register class to use for source operand given VT
1480 class getVregSrcForVT<ValueType VT> {
1481 RegisterClass ret = !if(!eq(VT.Size, 128), VReg_128,
1482 !if(!eq(VT.Size, 96), VReg_96,
1483 !if(!eq(VT.Size, 64), VReg_64,
1484 !if(!eq(VT.Size, 48), VReg_64,
1488 class getSDWASrcForVT <ValueType VT> {
1489 bit isFP = getIsFP<VT>.ret;
1490 RegisterOperand retFlt = !if(!eq(VT.Size, 16), SDWASrc_f16, SDWASrc_f32);
1491 RegisterOperand retInt = !if(!eq(VT.Size, 16), SDWASrc_i16, SDWASrc_i32);
1492 RegisterOperand ret = !if(isFP, retFlt, retInt);
1495 // Returns the register class to use for sources of VOP3 instructions for the
1497 class getVOP3SrcForVT<ValueType VT> {
1498 bit isFP = getIsFP<VT>.ret;
1499 RegisterOperand ret =
1500 !if(!eq(VT.Size, 128),
1502 !if(!eq(VT.Size, 64),
1506 !if(!eq(VT.Value, i1.Value),
1509 !if(!eq(VT.Value, f16.Value),
1511 !if(!eq(VT.Value, v2f16.Value),
1513 !if(!eq(VT.Value, v4f16.Value),
1519 !if(!eq(VT.Value, i16.Value),
1521 !if(!eq(VT.Value, v2i16.Value),
1532 // Float or packed int
1533 class isModifierType<ValueType SrcVT> {
1535 !if(!eq(SrcVT.Value, f16.Value), 1,
1536 !if(!eq(SrcVT.Value, f32.Value), 1,
1537 !if(!eq(SrcVT.Value, f64.Value), 1,
1538 !if(!eq(SrcVT.Value, v2f16.Value), 1,
1539 !if(!eq(SrcVT.Value, v2i16.Value), 1,
1543 // Return type of input modifiers operand for specified input operand
1544 class getSrcMod <ValueType VT, bit EnableF32SrcMods> {
1545 bit isFP = getIsFP<VT>.ret;
1546 bit isPacked = isPackedType<VT>.ret;
1547 Operand ret = !if(!eq(VT.Size, 64),
1548 !if(isFP, FP64InputMods, Int64InputMods),
1550 !if(!eq(VT.Value, f16.Value),
1554 !if(EnableF32SrcMods, FP32InputMods, Int32InputMods))
1558 class getOpSelMod <ValueType VT> {
1559 Operand ret = !if(!eq(VT.Value, f16.Value), FP16InputMods, IntOpSelMods);
1562 // Return type of input modifiers operand specified input operand for DPP
1563 class getSrcModExt <ValueType VT> {
1564 bit isFP = getIsFP<VT>.ret;
1565 Operand ret = !if(isFP, FPVRegInputMods, IntVRegInputMods);
1568 // Return type of input modifiers operand specified input operand for SDWA
1569 class getSrcModSDWA <ValueType VT> {
1570 Operand ret = !if(!eq(VT.Value, f16.Value), FP16SDWAInputMods,
1571 !if(!eq(VT.Value, f32.Value), FP32SDWAInputMods,
1572 !if(!eq(VT.Value, i16.Value), Int16SDWAInputMods,
1573 Int32SDWAInputMods)));
1576 // Returns the input arguments for VOP[12C] instructions for the given SrcVT.
1577 class getIns32 <RegisterOperand Src0RC, RegisterClass Src1RC, int NumSrcArgs> {
1578 dag ret = !if(!eq(NumSrcArgs, 1), (ins Src0RC:$src0), // VOP1
1579 !if(!eq(NumSrcArgs, 2), (ins Src0RC:$src0, Src1RC:$src1), // VOP2
1583 // Returns the input arguments for VOP3 instructions for the given SrcVT.
1584 class getIns64 <RegisterOperand Src0RC, RegisterOperand Src1RC,
1585 RegisterOperand Src2RC, int NumSrcArgs,
1586 bit HasIntClamp, bit HasModifiers, bit HasSrc2Mods, bit HasOMod,
1587 Operand Src0Mod, Operand Src1Mod, Operand Src2Mod> {
1590 !if (!eq(NumSrcArgs, 0),
1591 // VOP1 without input operands (V_NOP, V_CLREXCP)
1594 !if (!eq(NumSrcArgs, 1),
1595 !if (!eq(HasModifiers, 1),
1596 // VOP1 with modifiers
1597 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1598 clampmod:$clamp, omod:$omod)
1600 // VOP1 without modifiers
1601 !if (!eq(HasIntClamp, 1),
1602 (ins Src0RC:$src0, clampmod:$clamp),
1605 !if (!eq(NumSrcArgs, 2),
1606 !if (!eq(HasModifiers, 1),
1607 // VOP 2 with modifiers
1608 !if( !eq(HasOMod, 1),
1609 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1610 Src1Mod:$src1_modifiers, Src1RC:$src1,
1611 clampmod:$clamp, omod:$omod),
1612 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1613 Src1Mod:$src1_modifiers, Src1RC:$src1,
1616 // VOP2 without modifiers
1617 !if (!eq(HasIntClamp, 1),
1618 (ins Src0RC:$src0, Src1RC:$src1, clampmod:$clamp),
1619 (ins Src0RC:$src0, Src1RC:$src1))
1622 /* NumSrcArgs == 3 */,
1623 !if (!eq(HasModifiers, 1),
1624 !if (!eq(HasSrc2Mods, 1),
1625 // VOP3 with modifiers
1626 !if (!eq(HasOMod, 1),
1627 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1628 Src1Mod:$src1_modifiers, Src1RC:$src1,
1629 Src2Mod:$src2_modifiers, Src2RC:$src2,
1630 clampmod:$clamp, omod:$omod),
1631 !if (!eq(HasIntClamp, 1),
1632 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1633 Src1Mod:$src1_modifiers, Src1RC:$src1,
1634 Src2Mod:$src2_modifiers, Src2RC:$src2,
1636 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1637 Src1Mod:$src1_modifiers, Src1RC:$src1,
1638 Src2Mod:$src2_modifiers, Src2RC:$src2))),
1639 // VOP3 with modifiers except src2
1640 !if (!eq(HasOMod, 1),
1641 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1642 Src1Mod:$src1_modifiers, Src1RC:$src1,
1643 Src2RC:$src2, clampmod:$clamp, omod:$omod),
1644 !if (!eq(HasIntClamp, 1),
1645 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1646 Src1Mod:$src1_modifiers, Src1RC:$src1,
1647 Src2RC:$src2, clampmod:$clamp),
1648 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1649 Src1Mod:$src1_modifiers, Src1RC:$src1,
1652 // VOP3 without modifiers
1653 !if (!eq(HasIntClamp, 1),
1654 (ins Src0RC:$src0, Src1RC:$src1, Src2RC:$src2, clampmod:$clamp),
1655 (ins Src0RC:$src0, Src1RC:$src1, Src2RC:$src2))
1659 /// XXX - src1 may only allow VGPRs?
1661 // The modifiers (except clamp) are dummy operands for the benefit of
1662 // printing and parsing. They defer their values to looking at the
1663 // srcN_modifiers for what to print.
1664 class getInsVOP3P <RegisterOperand Src0RC, RegisterOperand Src1RC,
1665 RegisterOperand Src2RC, int NumSrcArgs,
1667 Operand Src0Mod, Operand Src1Mod, Operand Src2Mod> {
1668 dag ret = !if (!eq(NumSrcArgs, 2),
1670 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1671 Src1Mod:$src1_modifiers, Src1RC:$src1,
1673 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1674 neg_lo:$neg_lo, neg_hi:$neg_hi),
1675 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1676 Src1Mod:$src1_modifiers, Src1RC:$src1,
1677 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1678 neg_lo:$neg_lo, neg_hi:$neg_hi)),
1679 // else NumSrcArgs == 3
1681 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1682 Src1Mod:$src1_modifiers, Src1RC:$src1,
1683 Src2Mod:$src2_modifiers, Src2RC:$src2,
1685 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1686 neg_lo:$neg_lo, neg_hi:$neg_hi),
1687 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1688 Src1Mod:$src1_modifiers, Src1RC:$src1,
1689 Src2Mod:$src2_modifiers, Src2RC:$src2,
1690 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1691 neg_lo:$neg_lo, neg_hi:$neg_hi))
1695 class getInsVOP3OpSel <RegisterOperand Src0RC,
1696 RegisterOperand Src1RC,
1697 RegisterOperand Src2RC,
1703 dag ret = !if (!eq(NumSrcArgs, 2),
1705 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1706 Src1Mod:$src1_modifiers, Src1RC:$src1,
1709 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1710 Src1Mod:$src1_modifiers, Src1RC:$src1,
1712 // else NumSrcArgs == 3
1714 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1715 Src1Mod:$src1_modifiers, Src1RC:$src1,
1716 Src2Mod:$src2_modifiers, Src2RC:$src2,
1719 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1720 Src1Mod:$src1_modifiers, Src1RC:$src1,
1721 Src2Mod:$src2_modifiers, Src2RC:$src2,
1726 class getInsDPP <RegisterOperand DstRC, RegisterClass Src0RC, RegisterClass Src1RC,
1727 int NumSrcArgs, bit HasModifiers,
1728 Operand Src0Mod, Operand Src1Mod> {
1730 dag ret = !if (!eq(NumSrcArgs, 0),
1731 // VOP1 without input operands (V_NOP)
1732 (ins dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1733 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl),
1734 !if (!eq(NumSrcArgs, 1),
1735 !if (!eq(HasModifiers, 1),
1736 // VOP1_DPP with modifiers
1737 (ins DstRC:$old, Src0Mod:$src0_modifiers,
1738 Src0RC:$src0, dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1739 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl)
1741 // VOP1_DPP without modifiers
1742 (ins DstRC:$old, Src0RC:$src0,
1743 dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1744 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl)
1746 /* NumSrcArgs == 2 */,
1747 !if (!eq(HasModifiers, 1),
1748 // VOP2_DPP with modifiers
1750 Src0Mod:$src0_modifiers, Src0RC:$src0,
1751 Src1Mod:$src1_modifiers, Src1RC:$src1,
1752 dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1753 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl)
1755 // VOP2_DPP without modifiers
1757 Src0RC:$src0, Src1RC:$src1, dpp_ctrl:$dpp_ctrl,
1758 row_mask:$row_mask, bank_mask:$bank_mask,
1759 bound_ctrl:$bound_ctrl)
1763 class getInsDPP16 <RegisterOperand DstRC, RegisterClass Src0RC, RegisterClass Src1RC,
1764 int NumSrcArgs, bit HasModifiers,
1765 Operand Src0Mod, Operand Src1Mod> {
1766 dag ret = !con(getInsDPP<DstRC, Src0RC, Src1RC, NumSrcArgs,
1767 HasModifiers, Src0Mod, Src1Mod>.ret,
1771 class getInsDPP8 <RegisterOperand DstRC, RegisterClass Src0RC, RegisterClass Src1RC,
1772 int NumSrcArgs, bit HasModifiers,
1773 Operand Src0Mod, Operand Src1Mod> {
1774 dag ret = !if (!eq(NumSrcArgs, 0),
1775 // VOP1 without input operands (V_NOP)
1776 (ins dpp8:$dpp8, FI:$fi),
1777 !if (!eq(NumSrcArgs, 1),
1778 !if (!eq(HasModifiers, 1),
1779 // VOP1_DPP with modifiers
1780 (ins DstRC:$old, Src0Mod:$src0_modifiers,
1781 Src0RC:$src0, dpp8:$dpp8, FI:$fi)
1783 // VOP1_DPP without modifiers
1784 (ins DstRC:$old, Src0RC:$src0, dpp8:$dpp8, FI:$fi)
1786 /* NumSrcArgs == 2 */,
1787 !if (!eq(HasModifiers, 1),
1788 // VOP2_DPP with modifiers
1790 Src0Mod:$src0_modifiers, Src0RC:$src0,
1791 Src1Mod:$src1_modifiers, Src1RC:$src1,
1794 // VOP2_DPP without modifiers
1796 Src0RC:$src0, Src1RC:$src1, dpp8:$dpp8, FI:$fi)
1802 class getInsSDWA <RegisterOperand Src0RC, RegisterOperand Src1RC, int NumSrcArgs,
1803 bit HasSDWAOMod, Operand Src0Mod, Operand Src1Mod,
1806 dag ret = !if(!eq(NumSrcArgs, 0),
1807 // VOP1 without input operands (V_NOP)
1809 !if(!eq(NumSrcArgs, 1),
1811 !if(!eq(HasSDWAOMod, 0),
1812 // VOP1_SDWA without omod
1813 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1815 dst_sel:$dst_sel, dst_unused:$dst_unused,
1816 src0_sel:$src0_sel),
1817 // VOP1_SDWA with omod
1818 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1819 clampmod:$clamp, omod:$omod,
1820 dst_sel:$dst_sel, dst_unused:$dst_unused,
1821 src0_sel:$src0_sel)),
1822 !if(!eq(NumSrcArgs, 2),
1823 !if(!eq(DstVT.Size, 1),
1825 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1826 Src1Mod:$src1_modifiers, Src1RC:$src1,
1827 clampmod:$clamp, src0_sel:$src0_sel, src1_sel:$src1_sel),
1829 !if(!eq(HasSDWAOMod, 0),
1830 // VOP2_SDWA without omod
1831 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1832 Src1Mod:$src1_modifiers, Src1RC:$src1,
1834 dst_sel:$dst_sel, dst_unused:$dst_unused,
1835 src0_sel:$src0_sel, src1_sel:$src1_sel),
1836 // VOP2_SDWA with omod
1837 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1838 Src1Mod:$src1_modifiers, Src1RC:$src1,
1839 clampmod:$clamp, omod:$omod,
1840 dst_sel:$dst_sel, dst_unused:$dst_unused,
1841 src0_sel:$src0_sel, src1_sel:$src1_sel))),
1842 (ins)/* endif */)));
1845 // Outs for DPP and SDWA
1846 class getOutsExt <bit HasDst, ValueType DstVT, RegisterOperand DstRCExt> {
1847 dag ret = !if(HasDst,
1848 !if(!eq(DstVT.Size, 1),
1849 (outs), // no dst for VOPC, we use "vcc"-token as dst in SDWA VOPC instructions
1850 (outs DstRCExt:$vdst)),
1855 class getOutsSDWA <bit HasDst, ValueType DstVT, RegisterOperand DstRCSDWA> {
1856 dag ret = !if(HasDst,
1857 !if(!eq(DstVT.Size, 1),
1858 (outs DstRCSDWA:$sdst),
1859 (outs DstRCSDWA:$vdst)),
1863 // Returns the assembly string for the inputs and outputs of a VOP[12C]
1864 // instruction. This does not add the _e32 suffix, so it can be reused
1866 class getAsm32 <bit HasDst, int NumSrcArgs, ValueType DstVT = i32> {
1867 string dst = !if(!eq(DstVT.Size, 1), "$sdst", "$vdst"); // use $sdst for VOPC
1868 string src0 = ", $src0";
1869 string src1 = ", $src1";
1870 string src2 = ", $src2";
1871 string ret = !if(HasDst, dst, "") #
1872 !if(!eq(NumSrcArgs, 1), src0, "") #
1873 !if(!eq(NumSrcArgs, 2), src0#src1, "") #
1874 !if(!eq(NumSrcArgs, 3), src0#src1#src2, "");
1877 // Returns the assembly string for the inputs and outputs of a VOP3
1879 class getAsm64 <bit HasDst, int NumSrcArgs, bit HasIntClamp, bit HasModifiers,
1880 bit HasOMod, ValueType DstVT = i32> {
1881 string dst = !if(!eq(DstVT.Size, 1), "$sdst", "$vdst"); // use $sdst for VOPC
1882 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1883 string src1 = !if(!eq(NumSrcArgs, 1), "",
1884 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1885 " $src1_modifiers,"));
1886 string src2 = !if(!eq(NumSrcArgs, 3), " $src2_modifiers", "");
1887 string iclamp = !if(HasIntClamp, "$clamp", "");
1889 !if(!eq(HasModifiers, 0),
1890 getAsm32<HasDst, NumSrcArgs, DstVT>.ret # iclamp,
1891 dst#", "#src0#src1#src2#"$clamp"#!if(HasOMod, "$omod", ""));
1894 // Returns the assembly string for the inputs and outputs of a VOP3P
1896 class getAsmVOP3P <bit HasDst, int NumSrcArgs, bit HasModifiers,
1897 bit HasClamp, ValueType DstVT = i32> {
1898 string dst = " $vdst";
1899 string src0 = !if(!eq(NumSrcArgs, 1), "$src0", "$src0,");
1900 string src1 = !if(!eq(NumSrcArgs, 1), "",
1901 !if(!eq(NumSrcArgs, 2), " $src1",
1903 string src2 = !if(!eq(NumSrcArgs, 3), " $src2", "");
1905 string mods = !if(HasModifiers, "$neg_lo$neg_hi", "");
1906 string clamp = !if(HasClamp, "$clamp", "");
1908 // Each modifier is printed as an array of bits for each operand, so
1909 // all operands are printed as part of src0_modifiers.
1910 string ret = dst#", "#src0#src1#src2#"$op_sel$op_sel_hi"#mods#clamp;
1913 class getAsmVOP3OpSel <int NumSrcArgs,
1918 string dst = " $vdst";
1920 string isrc0 = !if(!eq(NumSrcArgs, 1), "$src0", "$src0,");
1921 string isrc1 = !if(!eq(NumSrcArgs, 1), "",
1922 !if(!eq(NumSrcArgs, 2), " $src1",
1924 string isrc2 = !if(!eq(NumSrcArgs, 3), " $src2", "");
1926 string fsrc0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1927 string fsrc1 = !if(!eq(NumSrcArgs, 1), "",
1928 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1929 " $src1_modifiers,"));
1930 string fsrc2 = !if(!eq(NumSrcArgs, 3), " $src2_modifiers", "");
1932 string src0 = !if(Src0HasMods, fsrc0, isrc0);
1933 string src1 = !if(Src1HasMods, fsrc1, isrc1);
1934 string src2 = !if(Src2HasMods, fsrc2, isrc2);
1936 string clamp = !if(HasClamp, "$clamp", "");
1938 string ret = dst#", "#src0#src1#src2#"$op_sel"#clamp;
1941 class getAsmDPP <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> {
1942 string dst = !if(HasDst,
1943 !if(!eq(DstVT.Size, 1),
1946 ""); // use $sdst for VOPC
1947 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1948 string src1 = !if(!eq(NumSrcArgs, 1), "",
1949 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1950 " $src1_modifiers,"));
1951 string args = !if(!eq(HasModifiers, 0),
1952 getAsm32<0, NumSrcArgs, DstVT>.ret,
1954 string ret = dst#args#" $dpp_ctrl$row_mask$bank_mask$bound_ctrl";
1957 class getAsmDPP16 <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> {
1958 string ret = getAsmDPP<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret#"$fi";
1961 class getAsmDPP8 <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> {
1962 string dst = !if(HasDst,
1963 !if(!eq(DstVT.Size, 1),
1966 ""); // use $sdst for VOPC
1967 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1968 string src1 = !if(!eq(NumSrcArgs, 1), "",
1969 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1970 " $src1_modifiers,"));
1971 string args = !if(!eq(HasModifiers, 0),
1972 getAsm32<0, NumSrcArgs, DstVT>.ret,
1974 string ret = dst#args#"$dpp8$fi";
1977 class getAsmSDWA <bit HasDst, int NumSrcArgs, ValueType DstVT = i32> {
1978 string dst = !if(HasDst,
1979 !if(!eq(DstVT.Size, 1),
1980 " vcc", // use vcc token as dst for VOPC instructioins
1983 string src0 = "$src0_modifiers";
1984 string src1 = "$src1_modifiers";
1985 string args = !if(!eq(NumSrcArgs, 0),
1987 !if(!eq(NumSrcArgs, 1),
1989 ", "#src0#", "#src1#"$clamp"
1992 string sdwa = !if(!eq(NumSrcArgs, 0),
1994 !if(!eq(NumSrcArgs, 1),
1995 " $dst_sel $dst_unused $src0_sel",
1996 !if(!eq(DstVT.Size, 1),
1997 " $src0_sel $src1_sel", // No dst_sel and dst_unused for VOPC
1998 " $dst_sel $dst_unused $src0_sel $src1_sel"
2002 string ret = dst#args#sdwa;
2005 class getAsmSDWA9 <bit HasDst, bit HasOMod, int NumSrcArgs,
2006 ValueType DstVT = i32> {
2007 string dst = !if(HasDst,
2008 !if(!eq(DstVT.Size, 1),
2012 string src0 = "$src0_modifiers";
2013 string src1 = "$src1_modifiers";
2014 string out_mods = !if(!eq(HasOMod, 0), "$clamp", "$clamp$omod");
2015 string args = !if(!eq(NumSrcArgs, 0), "",
2016 !if(!eq(NumSrcArgs, 1),
2021 string sdwa = !if(!eq(NumSrcArgs, 0), "",
2022 !if(!eq(NumSrcArgs, 1),
2023 out_mods#" $dst_sel $dst_unused $src0_sel",
2024 !if(!eq(DstVT.Size, 1),
2025 " $src0_sel $src1_sel", // No dst_sel, dst_unused and output modifiers for VOPC
2026 out_mods#" $dst_sel $dst_unused $src0_sel $src1_sel"
2030 string ret = dst#args#sdwa;
2034 // Function that checks if instruction supports DPP and SDWA
2035 class getHasExt <int NumSrcArgs, ValueType DstVT = i32, ValueType Src0VT = i32,
2036 ValueType Src1VT = i32> {
2037 bit ret = !if(!eq(NumSrcArgs, 3),
2038 0, // NumSrcArgs == 3 - No DPP or SDWA for VOP3
2039 !if(!eq(DstVT.Size, 64),
2040 0, // 64-bit dst - No DPP or SDWA for 64-bit operands
2041 !if(!eq(Src0VT.Size, 64),
2043 !if(!eq(Src1VT.Size, 64),
2052 class getHasDPP <int NumSrcArgs, ValueType DstVT = i32, ValueType Src0VT = i32,
2053 ValueType Src1VT = i32> {
2054 bit ret = !if(!eq(NumSrcArgs, 0), 0,
2055 getHasExt<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret);
2058 class BitOr<bit a, bit b> {
2059 bit ret = !if(a, 1, !if(b, 1, 0));
2062 class BitAnd<bit a, bit b> {
2063 bit ret = !if(a, !if(b, 1, 0), 0);
2071 class VOPProfile <list<ValueType> _ArgVT, bit _EnableF32SrcMods = 0,
2072 bit _EnableClamp = 0> {
2074 field list<ValueType> ArgVT = _ArgVT;
2075 field bit EnableF32SrcMods = _EnableF32SrcMods;
2076 field bit EnableClamp = _EnableClamp;
2078 field ValueType DstVT = ArgVT[0];
2079 field ValueType Src0VT = ArgVT[1];
2080 field ValueType Src1VT = ArgVT[2];
2081 field ValueType Src2VT = ArgVT[3];
2082 field RegisterOperand DstRC = getVALUDstForVT<DstVT>.ret;
2083 field RegisterOperand DstRCDPP = getVALUDstForVT<DstVT>.ret;
2084 field RegisterOperand DstRCSDWA = getSDWADstForVT<DstVT>.ret;
2085 field RegisterOperand Src0RC32 = getVOPSrc0ForVT<Src0VT>.ret;
2086 field RegisterClass Src1RC32 = getVregSrcForVT<Src1VT>.ret;
2087 field RegisterOperand Src0RC64 = getVOP3SrcForVT<Src0VT>.ret;
2088 field RegisterOperand Src1RC64 = getVOP3SrcForVT<Src1VT>.ret;
2089 field RegisterOperand Src2RC64 = getVOP3SrcForVT<Src2VT>.ret;
2090 field RegisterClass Src0DPP = getVregSrcForVT<Src0VT>.ret;
2091 field RegisterClass Src1DPP = getVregSrcForVT<Src1VT>.ret;
2092 field RegisterOperand Src0SDWA = getSDWASrcForVT<Src0VT>.ret;
2093 field RegisterOperand Src1SDWA = getSDWASrcForVT<Src0VT>.ret;
2094 field Operand Src0Mod = getSrcMod<Src0VT, EnableF32SrcMods>.ret;
2095 field Operand Src1Mod = getSrcMod<Src1VT, EnableF32SrcMods>.ret;
2096 field Operand Src2Mod = getSrcMod<Src2VT, EnableF32SrcMods>.ret;
2097 field Operand Src0ModDPP = getSrcModExt<Src0VT>.ret;
2098 field Operand Src1ModDPP = getSrcModExt<Src1VT>.ret;
2099 field Operand Src0ModSDWA = getSrcModSDWA<Src0VT>.ret;
2100 field Operand Src1ModSDWA = getSrcModSDWA<Src1VT>.ret;
2103 field bit HasDst = !if(!eq(DstVT.Value, untyped.Value), 0, 1);
2104 field bit HasDst32 = HasDst;
2105 field bit EmitDst = HasDst; // force dst encoding, see v_movreld_b32 special case
2106 field int NumSrcArgs = getNumSrcArgs<Src0VT, Src1VT, Src2VT>.ret;
2107 field bit HasSrc0 = !if(!eq(Src0VT.Value, untyped.Value), 0, 1);
2108 field bit HasSrc1 = !if(!eq(Src1VT.Value, untyped.Value), 0, 1);
2109 field bit HasSrc2 = !if(!eq(Src2VT.Value, untyped.Value), 0, 1);
2111 // TODO: Modifiers logic is somewhat adhoc here, to be refined later
2112 // HasModifiers affects the normal and DPP encodings. We take note of EnableF32SrcMods, which
2113 // enables modifiers for i32 type.
2114 field bit HasModifiers = BitOr<isModifierType<Src0VT>.ret, EnableF32SrcMods>.ret;
2116 // HasSrc*FloatMods affects the SDWA encoding. We ignore EnableF32SrcMods.
2117 field bit HasSrc0FloatMods = isFloatType<Src0VT>.ret;
2118 field bit HasSrc1FloatMods = isFloatType<Src1VT>.ret;
2119 field bit HasSrc2FloatMods = isFloatType<Src2VT>.ret;
2121 // HasSrc*IntMods affects the SDWA encoding. We ignore EnableF32SrcMods.
2122 field bit HasSrc0IntMods = isIntType<Src0VT>.ret;
2123 field bit HasSrc1IntMods = isIntType<Src1VT>.ret;
2124 field bit HasSrc2IntMods = isIntType<Src2VT>.ret;
2126 field bit HasSrc0Mods = HasModifiers;
2127 field bit HasSrc1Mods = !if(HasModifiers, BitOr<HasSrc1FloatMods, HasSrc1IntMods>.ret, 0);
2128 field bit HasSrc2Mods = !if(HasModifiers, BitOr<HasSrc2FloatMods, HasSrc2IntMods>.ret, 0);
2130 field bit HasClamp = BitOr<isModifierType<Src0VT>.ret, EnableClamp>.ret;
2131 field bit HasSDWAClamp = EmitDst;
2132 field bit HasFPClamp = BitAnd<isFloatType<DstVT>.ret, HasClamp>.ret;
2133 field bit HasIntClamp = !if(isFloatType<DstVT>.ret, 0, HasClamp);
2134 field bit HasClampLo = HasClamp;
2135 field bit HasClampHi = BitAnd<isPackedType<DstVT>.ret, HasClamp>.ret;
2136 field bit HasHigh = 0;
2138 field bit IsPacked = isPackedType<Src0VT>.ret;
2139 field bit HasOpSel = IsPacked;
2140 field bit HasOMod = !if(HasOpSel, 0, isFloatType<DstVT>.ret);
2141 field bit HasSDWAOMod = isFloatType<DstVT>.ret;
2143 field bit HasExt = getHasExt<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret;
2144 field bit HasExtDPP = getHasDPP<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret;
2145 field bit HasExtSDWA = HasExt;
2146 field bit HasExtSDWA9 = HasExt;
2147 field int NeedPatGen = PatGenMode.NoPattern;
2149 field bit IsMAI = 0;
2150 field bit IsDOT = 0;
2152 field Operand Src0PackedMod = !if(HasSrc0FloatMods, PackedF16InputMods, PackedI16InputMods);
2153 field Operand Src1PackedMod = !if(HasSrc1FloatMods, PackedF16InputMods, PackedI16InputMods);
2154 field Operand Src2PackedMod = !if(HasSrc2FloatMods, PackedF16InputMods, PackedI16InputMods);
2156 field dag Outs = !if(HasDst,(outs DstRC:$vdst),(outs));
2158 // VOP3b instructions are a special case with a second explicit
2159 // output. This is manually overridden for them.
2160 field dag Outs32 = Outs;
2161 field dag Outs64 = Outs;
2162 field dag OutsDPP = getOutsExt<HasDst, DstVT, DstRCDPP>.ret;
2163 field dag OutsDPP8 = getOutsExt<HasDst, DstVT, DstRCDPP>.ret;
2164 field dag OutsSDWA = getOutsSDWA<HasDst, DstVT, DstRCSDWA>.ret;
2166 field dag Ins32 = getIns32<Src0RC32, Src1RC32, NumSrcArgs>.ret;
2167 field dag Ins64 = getIns64<Src0RC64, Src1RC64, Src2RC64, NumSrcArgs,
2168 HasIntClamp, HasModifiers, HasSrc2Mods,
2169 HasOMod, Src0Mod, Src1Mod, Src2Mod>.ret;
2170 field dag InsVOP3P = getInsVOP3P<Src0RC64, Src1RC64, Src2RC64,
2171 NumSrcArgs, HasClamp,
2172 Src0PackedMod, Src1PackedMod, Src2PackedMod>.ret;
2173 field dag InsVOP3OpSel = getInsVOP3OpSel<Src0RC64, Src1RC64, Src2RC64,
2176 getOpSelMod<Src0VT>.ret,
2177 getOpSelMod<Src1VT>.ret,
2178 getOpSelMod<Src2VT>.ret>.ret;
2179 field dag InsDPP = !if(HasExtDPP,
2180 getInsDPP<DstRCDPP, Src0DPP, Src1DPP, NumSrcArgs,
2181 HasModifiers, Src0ModDPP, Src1ModDPP>.ret,
2183 field dag InsDPP16 = getInsDPP16<DstRCDPP, Src0DPP, Src1DPP, NumSrcArgs,
2184 HasModifiers, Src0ModDPP, Src1ModDPP>.ret;
2185 field dag InsDPP8 = getInsDPP8<DstRCDPP, Src0DPP, Src1DPP, NumSrcArgs, 0,
2186 Src0ModDPP, Src1ModDPP>.ret;
2187 field dag InsSDWA = getInsSDWA<Src0SDWA, Src1SDWA, NumSrcArgs,
2188 HasSDWAOMod, Src0ModSDWA, Src1ModSDWA,
2192 field string Asm32 = getAsm32<HasDst, NumSrcArgs, DstVT>.ret;
2193 field string Asm64 = getAsm64<HasDst, NumSrcArgs, HasIntClamp, HasModifiers, HasOMod, DstVT>.ret;
2194 field string AsmVOP3P = getAsmVOP3P<HasDst, NumSrcArgs, HasModifiers, HasClamp, DstVT>.ret;
2195 field string AsmVOP3OpSel = getAsmVOP3OpSel<NumSrcArgs,
2199 HasSrc2FloatMods>.ret;
2200 field string AsmDPP = !if(HasExtDPP,
2201 getAsmDPP<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret, "");
2202 field string AsmDPP16 = getAsmDPP16<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret;
2203 field string AsmDPP8 = getAsmDPP8<HasDst, NumSrcArgs, 0, DstVT>.ret;
2204 field string AsmSDWA = getAsmSDWA<HasDst, NumSrcArgs, DstVT>.ret;
2205 field string AsmSDWA9 = getAsmSDWA9<HasDst, HasSDWAOMod, NumSrcArgs, DstVT>.ret;
2207 field string TieRegDPP = "$old";
2210 class VOP_NO_EXT <VOPProfile p> : VOPProfile <p.ArgVT> {
2214 let HasExtSDWA9 = 0;
2217 class VOP_PAT_GEN <VOPProfile p, int mode=PatGenMode.Pattern> : VOPProfile <p.ArgVT> {
2218 let NeedPatGen = mode;
2221 def VOP_F16_F16 : VOPProfile <[f16, f16, untyped, untyped]>;
2222 def VOP_F16_I16 : VOPProfile <[f16, i16, untyped, untyped]>;
2223 def VOP_I16_F16 : VOPProfile <[i16, f16, untyped, untyped]>;
2225 def VOP_F16_F16_F16 : VOPProfile <[f16, f16, f16, untyped]>;
2226 def VOP_F16_F16_I16 : VOPProfile <[f16, f16, i16, untyped]>;
2227 def VOP_F16_F16_I32 : VOPProfile <[f16, f16, i32, untyped]>;
2228 def VOP_I16_I16_I16 : VOPProfile <[i16, i16, i16, untyped]>;
2230 def VOP_I16_I16_I16_I16 : VOPProfile <[i16, i16, i16, i16, untyped]>;
2231 def VOP_F16_F16_F16_F16 : VOPProfile <[f16, f16, f16, f16, untyped]>;
2233 def VOP_I32_I16_I16_I32 : VOPProfile <[i32, i16, i16, i32, untyped]>;
2235 def VOP_V2F16_V2F16_V2F16 : VOPProfile <[v2f16, v2f16, v2f16, untyped]>;
2236 def VOP_V2I16_V2I16_V2I16 : VOPProfile <[v2i16, v2i16, v2i16, untyped]>;
2237 def VOP_B32_F16_F16 : VOPProfile <[i32, f16, f16, untyped]>;
2239 def VOP_V2F16_V2F16_V2F16_V2F16 : VOPProfile <[v2f16, v2f16, v2f16, v2f16]>;
2240 def VOP_V2I16_V2I16_V2I16_V2I16 : VOPProfile <[v2i16, v2i16, v2i16, v2i16]>;
2241 def VOP_V2I16_F32_F32 : VOPProfile <[v2i16, f32, f32, untyped]>;
2242 def VOP_V2I16_I32_I32 : VOPProfile <[v2i16, i32, i32, untyped]>;
2244 def VOP_F32_V2F16_V2F16_V2F16 : VOPProfile <[f32, v2f16, v2f16, v2f16]>;
2246 def VOP_NONE : VOPProfile <[untyped, untyped, untyped, untyped]>;
2248 def VOP_F32_F32 : VOPProfile <[f32, f32, untyped, untyped]>;
2249 def VOP_F32_F64 : VOPProfile <[f32, f64, untyped, untyped]>;
2250 def VOP_F32_I32 : VOPProfile <[f32, i32, untyped, untyped]>;
2251 def VOP_F64_F32 : VOPProfile <[f64, f32, untyped, untyped]>;
2252 def VOP_F64_F64 : VOPProfile <[f64, f64, untyped, untyped]>;
2253 def VOP_F64_I32 : VOPProfile <[f64, i32, untyped, untyped]>;
2254 def VOP_I32_F32 : VOPProfile <[i32, f32, untyped, untyped]>;
2255 def VOP_I32_F64 : VOPProfile <[i32, f64, untyped, untyped]>;
2256 def VOP_I32_I32 : VOPProfile <[i32, i32, untyped, untyped]>;
2257 def VOP_F16_F32 : VOPProfile <[f16, f32, untyped, untyped]>;
2258 def VOP_F32_F16 : VOPProfile <[f32, f16, untyped, untyped]>;
2260 def VOP_F32_F32_F16 : VOPProfile <[f32, f32, f16, untyped]>;
2261 def VOP_F32_F32_F32 : VOPProfile <[f32, f32, f32, untyped]>;
2262 def VOP_F32_F32_I32 : VOPProfile <[f32, f32, i32, untyped]>;
2263 def VOP_F64_F64_F64 : VOPProfile <[f64, f64, f64, untyped]>;
2264 def VOP_F64_F64_I32 : VOPProfile <[f64, f64, i32, untyped]>;
2265 def VOP_I32_F32_F32 : VOPProfile <[i32, f32, f32, untyped]>;
2266 def VOP_I32_F32_I32 : VOPProfile <[i32, f32, i32, untyped]>;
2267 def VOP_I32_I32_I32 : VOPProfile <[i32, i32, i32, untyped]>;
2268 def VOP_I32_I32_I32_ARITH : VOPProfile <[i32, i32, i32, untyped], 0, /*EnableClamp=*/1>;
2269 def VOP_V2F16_F32_F32 : VOPProfile <[v2f16, f32, f32, untyped]>;
2270 def VOP_F32_F16_F16_F16 : VOPProfile <[f32, f16, f16, f16]>;
2272 def VOP_I64_I64_I32 : VOPProfile <[i64, i64, i32, untyped]>;
2273 def VOP_I64_I32_I64 : VOPProfile <[i64, i32, i64, untyped]>;
2274 def VOP_I64_I64_I64 : VOPProfile <[i64, i64, i64, untyped]>;
2276 def VOP_F16_F32_F16_F32 : VOPProfile <[f16, f32, f16, f32]>;
2277 def VOP_F32_F32_F16_F16 : VOPProfile <[f32, f32, f16, f16]>;
2278 def VOP_F32_F32_F32_F32 : VOPProfile <[f32, f32, f32, f32]>;
2279 def VOP_F64_F64_F64_F64 : VOPProfile <[f64, f64, f64, f64]>;
2280 def VOP_I32_I32_I32_I32 : VOPProfile <[i32, i32, i32, i32]>;
2281 def VOP_I64_I32_I32_I64 : VOPProfile <[i64, i32, i32, i64]>;
2282 def VOP_I32_F32_I32_I32 : VOPProfile <[i32, f32, i32, i32]>;
2283 def VOP_I64_I64_I32_I64 : VOPProfile <[i64, i64, i32, i64]>;
2284 def VOP_V4I32_I64_I32_V4I32 : VOPProfile <[v4i32, i64, i32, v4i32]>;
2286 def VOP_F32_V2F16_V2F16_F32 : VOPProfile <[f32, v2f16, v2f16, f32]>;
2287 def VOP_I32_V2I16_V2I16_I32 : VOPProfile <[i32, v2i16, v2i16, i32]>;
2289 def VOP_V4F32_F32_F32_V4F32 : VOPProfile <[v4f32, f32, f32, v4f32]>;
2290 def VOP_V16F32_F32_F32_V16F32 : VOPProfile <[v16f32, f32, f32, v16f32]>;
2291 def VOP_V32F32_F32_F32_V32F32 : VOPProfile <[v32f32, f32, f32, v32f32]>;
2292 def VOP_V4F32_V4F16_V4F16_V4F32 : VOPProfile <[v4f32, v4f16, v4f16, v4f32]>;
2293 def VOP_V16F32_V4F16_V4F16_V16F32 : VOPProfile <[v16f32, v4f16, v4f16, v16f32]>;
2294 def VOP_V32F32_V4F16_V4F16_V32F32 : VOPProfile <[v32f32, v4f16, v4f16, v32f32]>;
2295 def VOP_V4F32_V2I16_V2I16_V4F32 : VOPProfile <[v4f32, v2i16, v2i16, v4f32]>;
2296 def VOP_V16F32_V2I16_V2I16_V16F32 : VOPProfile <[v16f32, v2i16, v2i16, v16f32]>;
2297 def VOP_V32F32_V2I16_V2I16_V32F32 : VOPProfile <[v32f32, v2i16, v2i16, v32f32]>;
2298 def VOP_V4I32_I32_I32_V4I32 : VOPProfile <[v4i32, i32, i32, v4i32]>;
2299 def VOP_V16I32_I32_I32_V16I32 : VOPProfile <[v16i32, i32, i32, v16i32]>;
2300 def VOP_V32I32_I32_I32_V32I32 : VOPProfile <[v32i32, i32, i32, v32i32]>;
2302 class Commutable_REV <string revOp, bit isOrig> {
2303 string RevOp = revOp;
2304 bit IsOrig = isOrig;
2307 class AtomicNoRet <string noRetOp, bit isRet> {
2308 string NoRetOp = noRetOp;
2312 //===----------------------------------------------------------------------===//
2313 // Interpolation opcodes
2314 //===----------------------------------------------------------------------===//
2316 class VINTRPDstOperand <RegisterClass rc> : RegisterOperand <rc, "printVINTRPDst">;
2318 class VINTRP_Pseudo <string opName, dag outs, dag ins, list<dag> pattern> :
2319 VINTRPCommon <outs, ins, "", pattern>,
2320 SIMCInstr<opName, SIEncodingFamily.NONE> {
2322 let isCodeGenOnly = 1;
2325 // FIXME-GFX10: WIP.
2326 class VINTRP_Real_si <bits <2> op, string opName, dag outs, dag ins,
2327 string asm, int encodingFamily> :
2328 VINTRPCommon <outs, ins, asm, []>,
2330 SIMCInstr<opName, encodingFamily> {
2331 let DisableDecoder = DisableSIDecoder;
2334 class VINTRP_Real_vi <bits <2> op, string opName, dag outs, dag ins,
2336 VINTRPCommon <outs, ins, asm, []>,
2338 SIMCInstr<opName, SIEncodingFamily.VI> {
2339 let AssemblerPredicate = VIAssemblerPredicate;
2340 let DecoderNamespace = "GFX8";
2341 let DisableDecoder = DisableVIDecoder;
2344 // FIXME-GFX10: WIP.
2345 multiclass VINTRP_m <bits <2> op, dag outs, dag ins, string asm,
2346 list<dag> pattern = []> {
2347 def "" : VINTRP_Pseudo <NAME, outs, ins, pattern>;
2349 let AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" in {
2350 def _si : VINTRP_Real_si <op, NAME, outs, ins, asm, SIEncodingFamily.SI>;
2351 } // End AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7"
2353 def _vi : VINTRP_Real_vi <op, NAME, outs, ins, asm>;
2355 let AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" in {
2356 def _gfx10 : VINTRP_Real_si<op, NAME, outs, ins, asm, SIEncodingFamily.GFX10>;
2357 } // End AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10"
2359 //===----------------------------------------------------------------------===//
2360 // Vector instruction mappings
2361 //===----------------------------------------------------------------------===//
2363 // Maps an opcode in e32 form to its e64 equivalent
2364 def getVOPe64 : InstrMapping {
2365 let FilterClass = "VOP";
2366 let RowFields = ["OpName"];
2367 let ColFields = ["Size", "VOP3"];
2368 let KeyCol = ["4", "0"];
2369 let ValueCols = [["8", "1"]];
2372 // Maps an opcode in e64 form to its e32 equivalent
2373 def getVOPe32 : InstrMapping {
2374 let FilterClass = "VOP";
2375 let RowFields = ["OpName"];
2376 let ColFields = ["Size", "VOP3"];
2377 let KeyCol = ["8", "1"];
2378 let ValueCols = [["4", "0"]];
2381 // Maps ordinary instructions to their SDWA counterparts
2382 def getSDWAOp : InstrMapping {
2383 let FilterClass = "VOP";
2384 let RowFields = ["OpName"];
2385 let ColFields = ["AsmVariantName"];
2386 let KeyCol = ["Default"];
2387 let ValueCols = [["SDWA"]];
2390 // Maps SDWA instructions to their ordinary counterparts
2391 def getBasicFromSDWAOp : InstrMapping {
2392 let FilterClass = "VOP";
2393 let RowFields = ["OpName"];
2394 let ColFields = ["AsmVariantName"];
2395 let KeyCol = ["SDWA"];
2396 let ValueCols = [["Default"]];
2399 // Maps ordinary instructions to their DPP counterparts
2400 def getDPPOp32 : InstrMapping {
2401 let FilterClass = "VOP";
2402 let RowFields = ["OpName"];
2403 let ColFields = ["AsmVariantName"];
2404 let KeyCol = ["Default"];
2405 let ValueCols = [["DPP"]];
2408 // Maps an commuted opcode to its original version
2409 def getCommuteOrig : InstrMapping {
2410 let FilterClass = "Commutable_REV";
2411 let RowFields = ["RevOp"];
2412 let ColFields = ["IsOrig"];
2414 let ValueCols = [["1"]];
2417 // Maps an original opcode to its commuted version
2418 def getCommuteRev : InstrMapping {
2419 let FilterClass = "Commutable_REV";
2420 let RowFields = ["RevOp"];
2421 let ColFields = ["IsOrig"];
2423 let ValueCols = [["0"]];
2426 def getMCOpcodeGen : InstrMapping {
2427 let FilterClass = "SIMCInstr";
2428 let RowFields = ["PseudoInstr"];
2429 let ColFields = ["Subtarget"];
2430 let KeyCol = [!cast<string>(SIEncodingFamily.NONE)];
2431 let ValueCols = [[!cast<string>(SIEncodingFamily.SI)],
2432 [!cast<string>(SIEncodingFamily.VI)],
2433 [!cast<string>(SIEncodingFamily.SDWA)],
2434 [!cast<string>(SIEncodingFamily.SDWA9)],
2435 // GFX80 encoding is added to work around a multiple matching
2436 // issue for buffer instructions with unpacked d16 data. This
2437 // does not actually change the encoding, and thus may be
2439 [!cast<string>(SIEncodingFamily.GFX80)],
2440 [!cast<string>(SIEncodingFamily.GFX9)],
2441 [!cast<string>(SIEncodingFamily.GFX10)],
2442 [!cast<string>(SIEncodingFamily.SDWA10)]];
2445 // Get equivalent SOPK instruction.
2446 def getSOPKOp : InstrMapping {
2447 let FilterClass = "SOPKInstTable";
2448 let RowFields = ["BaseCmpOp"];
2449 let ColFields = ["IsSOPK"];
2451 let ValueCols = [["1"]];
2454 def getAddr64Inst : InstrMapping {
2455 let FilterClass = "MUBUFAddr64Table";
2456 let RowFields = ["OpName"];
2457 let ColFields = ["IsAddr64"];
2459 let ValueCols = [["1"]];
2462 def getIfAddr64Inst : InstrMapping {
2463 let FilterClass = "MUBUFAddr64Table";
2464 let RowFields = ["OpName"];
2465 let ColFields = ["IsAddr64"];
2467 let ValueCols = [["1"]];
2470 def getMUBUFNoLdsInst : InstrMapping {
2471 let FilterClass = "MUBUFLdsTable";
2472 let RowFields = ["OpName"];
2473 let ColFields = ["IsLds"];
2475 let ValueCols = [["0"]];
2478 // Maps an atomic opcode to its version with a return value.
2479 def getAtomicRetOp : InstrMapping {
2480 let FilterClass = "AtomicNoRet";
2481 let RowFields = ["NoRetOp"];
2482 let ColFields = ["IsRet"];
2484 let ValueCols = [["1"]];
2487 // Maps an atomic opcode to its returnless version.
2488 def getAtomicNoRetOp : InstrMapping {
2489 let FilterClass = "AtomicNoRet";
2490 let RowFields = ["NoRetOp"];
2491 let ColFields = ["IsRet"];
2493 let ValueCols = [["0"]];
2496 // Maps a GLOBAL to its SADDR form.
2497 def getGlobalSaddrOp : InstrMapping {
2498 let FilterClass = "GlobalSaddrTable";
2499 let RowFields = ["SaddrOp"];
2500 let ColFields = ["IsSaddr"];
2502 let ValueCols = [["1"]];
2505 // Maps a v_cmpx opcode with sdst to opcode without sdst.
2506 def getVCMPXNoSDstOp : InstrMapping {
2507 let FilterClass = "VCMPXNoSDstTable";
2508 let RowFields = ["NoSDstOp"];
2509 let ColFields = ["HasSDst"];
2511 let ValueCols = [["0"]];
2514 // Maps a SOPP to a SOPP with S_NOP
2515 def getSOPPWithRelaxation : InstrMapping {
2516 let FilterClass = "Base_SOPP";
2517 let RowFields = ["AsmString"];
2518 let ColFields = ["Size"];
2520 let ValueCols = [["8"]];
2523 include "SIInstructions.td"
2525 include "DSInstructions.td"
2526 include "MIMGInstructions.td"