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 : LoadFrag <load_glue>, LocalAddress {
445 let MinAlignment = 8;
446 let IsNonExtLoad = 1;
448 def load_align16_local_m0 : LoadFrag <load_glue>, LocalAddress {
449 let MinAlignment = 16;
450 let IsNonExtLoad = 1;
455 def atomic_load_32_local_m0 : LoadFrag<atomic_load_32_glue>, LocalAddress;
456 def atomic_load_64_local_m0 : LoadFrag<atomic_load_64_glue>, LocalAddress;
459 def AMDGPUst_glue : SDNode <"ISD::STORE", SDTStore,
460 [SDNPHasChain, SDNPMayStore, SDNPMemOperand, SDNPInGlue]
463 def AMDGPUatomic_st_glue : SDNode <"ISD::ATOMIC_STORE", SDTAtomicStore,
464 [SDNPHasChain, SDNPMayStore, SDNPMemOperand, SDNPInGlue]
467 def atomic_store_glue : PatFrag<(ops node:$ptr, node:$val),
468 (AMDGPUatomic_st_glue node:$ptr, node:$val)> {
471 def unindexedstore_glue : PatFrag<(ops node:$val, node:$ptr),
472 (AMDGPUst_glue node:$val, node:$ptr)> {
477 def store_glue : PatFrag<(ops node:$val, node:$ptr),
478 (unindexedstore_glue node:$val, node:$ptr)> {
480 let IsTruncStore = 0;
483 def truncstore_glue : PatFrag<(ops node:$val, node:$ptr),
484 (unindexedstore_glue node:$val, node:$ptr)> {
486 let IsTruncStore = 1;
489 def truncstorei8_glue : PatFrag<(ops node:$val, node:$ptr),
490 (truncstore_glue node:$val, node:$ptr)> {
495 def truncstorei16_glue : PatFrag<(ops node:$val, node:$ptr),
496 (truncstore_glue node:$val, node:$ptr)> {
501 let IsStore = 1, AddressSpaces = StoreAddress_local.AddrSpaces in {
502 def store_glue_align8 : PatFrag<(ops node:$val, node:$ptr),
503 (store_glue node:$val, node:$ptr)>, Aligned<8>;
504 def store_glue_align16 : PatFrag<(ops node:$val, node:$ptr),
505 (store_glue node:$val, node:$ptr)>, Aligned<16>;
507 def store_local_m0 : PatFrag<(ops node:$val, node:$ptr),
508 (unindexedstore_glue node:$val, node:$ptr)> {
510 let IsTruncStore = 0;
513 def truncstorei8_local_m0 : PatFrag<(ops node:$val, node:$ptr),
514 (unindexedstore_glue node:$val, node:$ptr)> {
520 def truncstorei16_local_m0 : PatFrag<(ops node:$val, node:$ptr),
521 (unindexedstore_glue node:$val, node:$ptr)> {
526 // FIXME: atomic store doesn't work.
527 def atomic_store_local_m0 : StoreFrag<AMDGPUatomic_st_glue>, LocalAddress;
528 def store_align8_local_m0 : StoreFrag<store_glue_align8>, LocalAddress {
529 let IsTruncStore = 0;
532 def store_align16_local_m0 : StoreFrag<store_glue_align16>, LocalAddress {
533 let IsTruncStore = 0;
537 def si_setcc_uniform : PatFrag <
538 (ops node:$lhs, node:$rhs, node:$cond),
539 (setcc node:$lhs, node:$rhs, node:$cond), [{
540 for (SDNode *Use : N->uses()) {
541 if (Use->isMachineOpcode() || Use->getOpcode() != ISD::CopyToReg)
544 unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
545 if (Reg != AMDGPU::SCC)
551 //===----------------------------------------------------------------------===//
552 // SDNodes PatFrags for d16 loads
553 //===----------------------------------------------------------------------===//
555 class LoadD16Frag <SDPatternOperator op> : PatFrag<(ops node:$ptr, node:$tied_in), (op node:$ptr, node:$tied_in)>;
556 class LocalLoadD16 <SDPatternOperator op> : LoadD16Frag <op>, LocalAddress;
557 class GlobalLoadD16 <SDPatternOperator op> : LoadD16Frag <op>, GlobalLoadAddress;
558 class PrivateLoadD16 <SDPatternOperator op> : LoadD16Frag <op>, PrivateAddress;
559 class FlatLoadD16 <SDPatternOperator op> : LoadD16Frag <op>, FlatLoadAddress;
561 def load_d16_hi_local : LocalLoadD16 <SIload_d16_hi>;
562 def az_extloadi8_d16_hi_local : LocalLoadD16 <SIload_d16_hi_u8>;
563 def sextloadi8_d16_hi_local : LocalLoadD16 <SIload_d16_hi_i8>;
565 def load_d16_hi_global : GlobalLoadD16 <SIload_d16_hi>;
566 def az_extloadi8_d16_hi_global : GlobalLoadD16 <SIload_d16_hi_u8>;
567 def sextloadi8_d16_hi_global : GlobalLoadD16 <SIload_d16_hi_i8>;
569 def load_d16_hi_private : PrivateLoadD16 <SIload_d16_hi>;
570 def az_extloadi8_d16_hi_private : PrivateLoadD16 <SIload_d16_hi_u8>;
571 def sextloadi8_d16_hi_private : PrivateLoadD16 <SIload_d16_hi_i8>;
573 def load_d16_hi_flat : FlatLoadD16 <SIload_d16_hi>;
574 def az_extloadi8_d16_hi_flat : FlatLoadD16 <SIload_d16_hi_u8>;
575 def sextloadi8_d16_hi_flat : FlatLoadD16 <SIload_d16_hi_i8>;
578 def load_d16_lo_local : LocalLoadD16 <SIload_d16_lo>;
579 def az_extloadi8_d16_lo_local : LocalLoadD16 <SIload_d16_lo_u8>;
580 def sextloadi8_d16_lo_local : LocalLoadD16 <SIload_d16_lo_i8>;
582 def load_d16_lo_global : GlobalLoadD16 <SIload_d16_lo>;
583 def az_extloadi8_d16_lo_global : GlobalLoadD16 <SIload_d16_lo_u8>;
584 def sextloadi8_d16_lo_global : GlobalLoadD16 <SIload_d16_lo_i8>;
586 def load_d16_lo_private : PrivateLoadD16 <SIload_d16_lo>;
587 def az_extloadi8_d16_lo_private : PrivateLoadD16 <SIload_d16_lo_u8>;
588 def sextloadi8_d16_lo_private : PrivateLoadD16 <SIload_d16_lo_i8>;
590 def load_d16_lo_flat : FlatLoadD16 <SIload_d16_lo>;
591 def az_extloadi8_d16_lo_flat : FlatLoadD16 <SIload_d16_lo_u8>;
592 def sextloadi8_d16_lo_flat : FlatLoadD16 <SIload_d16_lo_i8>;
596 def lshr_rev : PatFrag <
597 (ops node:$src1, node:$src0),
601 def ashr_rev : PatFrag <
602 (ops node:$src1, node:$src0),
606 def lshl_rev : PatFrag <
607 (ops node:$src1, node:$src0),
611 multiclass SIAtomicM0Glue2 <string op_name, bit is_amdgpu = 0,
612 SDTypeProfile tc = SDTAtomic2,
616 !if(is_amdgpu, "AMDGPUISD", "ISD")#"::ATOMIC_"#op_name, tc,
617 [SDNPHasChain, SDNPMayStore, SDNPMayLoad, SDNPMemOperand, SDNPInGlue]
620 let AddressSpaces = StoreAddress_local.AddrSpaces in {
621 defm _local_m0 : binary_atomic_op <!cast<SDNode>(NAME#"_glue"), IsInt>;
624 let AddressSpaces = StoreAddress_region.AddrSpaces in {
625 defm _region_m0 : binary_atomic_op <!cast<SDNode>(NAME#"_glue"), IsInt>;
629 defm atomic_load_add : SIAtomicM0Glue2 <"LOAD_ADD">;
630 defm atomic_load_sub : SIAtomicM0Glue2 <"LOAD_SUB">;
631 defm atomic_inc : SIAtomicM0Glue2 <"INC", 1>;
632 defm atomic_dec : SIAtomicM0Glue2 <"DEC", 1>;
633 defm atomic_load_and : SIAtomicM0Glue2 <"LOAD_AND">;
634 defm atomic_load_min : SIAtomicM0Glue2 <"LOAD_MIN">;
635 defm atomic_load_max : SIAtomicM0Glue2 <"LOAD_MAX">;
636 defm atomic_load_or : SIAtomicM0Glue2 <"LOAD_OR">;
637 defm atomic_load_xor : SIAtomicM0Glue2 <"LOAD_XOR">;
638 defm atomic_load_umin : SIAtomicM0Glue2 <"LOAD_UMIN">;
639 defm atomic_load_umax : SIAtomicM0Glue2 <"LOAD_UMAX">;
640 defm atomic_swap : SIAtomicM0Glue2 <"SWAP">;
641 defm atomic_load_fadd : SIAtomicM0Glue2 <"LOAD_FADD", 0, SDTAtomic2_f32, 0>;
642 defm atomic_load_fmin : SIAtomicM0Glue2 <"LOAD_FMIN", 1, SDTAtomic2_f32, 0>;
643 defm atomic_load_fmax : SIAtomicM0Glue2 <"LOAD_FMAX", 1, SDTAtomic2_f32, 0>;
645 def as_i1imm : SDNodeXForm<imm, [{
646 return CurDAG->getTargetConstant(N->getZExtValue(), SDLoc(N), MVT::i1);
649 def as_i8imm : SDNodeXForm<imm, [{
650 return CurDAG->getTargetConstant(N->getZExtValue(), SDLoc(N), MVT::i8);
653 def as_i16imm : SDNodeXForm<imm, [{
654 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i16);
657 def as_i32imm: SDNodeXForm<imm, [{
658 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32);
661 def as_i64imm: SDNodeXForm<imm, [{
662 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i64);
665 def cond_as_i32imm: SDNodeXForm<cond, [{
666 return CurDAG->getTargetConstant(N->get(), SDLoc(N), MVT::i32);
669 // Copied from the AArch64 backend:
670 def bitcast_fpimm_to_i32 : SDNodeXForm<fpimm, [{
671 return CurDAG->getTargetConstant(
672 N->getValueAPF().bitcastToAPInt().getZExtValue(), SDLoc(N), MVT::i32);
675 def frameindex_to_targetframeindex : SDNodeXForm<frameindex, [{
676 auto FI = cast<FrameIndexSDNode>(N);
677 return CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
680 // Copied from the AArch64 backend:
681 def bitcast_fpimm_to_i64 : SDNodeXForm<fpimm, [{
682 return CurDAG->getTargetConstant(
683 N->getValueAPF().bitcastToAPInt().getZExtValue(), SDLoc(N), MVT::i64);
686 class bitextract_imm<int bitnum> : SDNodeXForm<imm, [{
687 uint64_t Imm = N->getZExtValue();
688 unsigned Bit = (Imm >> }] # bitnum # [{ ) & 1;
689 return CurDAG->getTargetConstant(Bit, SDLoc(N), MVT::i1);
692 def SIMM16bit : ImmLeaf <i32,
693 [{return isInt<16>(Imm);}]
696 def UIMM16bit : ImmLeaf <i32,
697 [{return isUInt<16>(Imm);}]
700 class InlineImm <ValueType vt> : PatLeaf <(vt imm), [{
701 return isInlineImmediate(N);
704 class InlineFPImm <ValueType vt> : PatLeaf <(vt fpimm), [{
705 return isInlineImmediate(N);
708 class VGPRImm <dag frag> : PatLeaf<frag, [{
712 def NegateImm : SDNodeXForm<imm, [{
713 return CurDAG->getConstant(-N->getSExtValue(), SDLoc(N), MVT::i32);
716 // TODO: When FP inline imm values work?
717 def NegSubInlineConst32 : ImmLeaf<i32, [{
718 return Imm < -16 && Imm >= -64;
721 def NegSubInlineConst16 : ImmLeaf<i16, [{
722 return Imm < -16 && Imm >= -64;
725 def ShiftAmt32Imm : PatLeaf <(imm), [{
726 return N->getZExtValue() < 32;
729 def getNegV2I16Imm : SDNodeXForm<build_vector, [{
730 return SDValue(packNegConstantV2I16(N, *CurDAG), 0);
733 def NegSubInlineConstV216 : PatLeaf<(build_vector), [{
734 assert(N->getNumOperands() == 2);
735 assert(N->getOperand(0).getValueType().getSizeInBits() == 16);
736 SDValue Src0 = N->getOperand(0);
737 SDValue Src1 = N->getOperand(1);
739 return isNegInlineImmediate(Src0.getNode());
741 return (isNullConstantOrUndef(Src0) && isNegInlineImmediate(Src1.getNode())) ||
742 (isNullConstantOrUndef(Src1) && isNegInlineImmediate(Src0.getNode()));
745 //===----------------------------------------------------------------------===//
747 //===----------------------------------------------------------------------===//
749 def SoppBrTarget : AsmOperandClass {
750 let Name = "SoppBrTarget";
751 let ParserMethod = "parseSOppBrTarget";
754 def sopp_brtarget : Operand<OtherVT> {
755 let EncoderMethod = "getSOPPBrEncoding";
756 let DecoderMethod = "decodeSoppBrTarget";
757 let OperandType = "OPERAND_PCREL";
758 let ParserMatchClass = SoppBrTarget;
761 def si_ga : Operand<iPTR>;
763 def InterpSlotMatchClass : AsmOperandClass {
764 let Name = "InterpSlot";
765 let PredicateMethod = "isInterpSlot";
766 let ParserMethod = "parseInterpSlot";
767 let RenderMethod = "addImmOperands";
770 def InterpSlot : Operand<i32> {
771 let PrintMethod = "printInterpSlot";
772 let ParserMatchClass = InterpSlotMatchClass;
773 let OperandType = "OPERAND_IMMEDIATE";
776 def AttrMatchClass : AsmOperandClass {
778 let PredicateMethod = "isInterpAttr";
779 let ParserMethod = "parseInterpAttr";
780 let RenderMethod = "addImmOperands";
783 // It appears to be necessary to create a separate operand for this to
784 // be able to parse attr<num> with no space.
785 def Attr : Operand<i32> {
786 let PrintMethod = "printInterpAttr";
787 let ParserMatchClass = AttrMatchClass;
788 let OperandType = "OPERAND_IMMEDIATE";
791 def AttrChanMatchClass : AsmOperandClass {
792 let Name = "AttrChan";
793 let PredicateMethod = "isAttrChan";
794 let RenderMethod = "addImmOperands";
797 def AttrChan : Operand<i32> {
798 let PrintMethod = "printInterpAttrChan";
799 let ParserMatchClass = AttrChanMatchClass;
800 let OperandType = "OPERAND_IMMEDIATE";
803 def SendMsgMatchClass : AsmOperandClass {
804 let Name = "SendMsg";
805 let PredicateMethod = "isSendMsg";
806 let ParserMethod = "parseSendMsgOp";
807 let RenderMethod = "addImmOperands";
810 def SwizzleMatchClass : AsmOperandClass {
811 let Name = "Swizzle";
812 let PredicateMethod = "isSwizzle";
813 let ParserMethod = "parseSwizzleOp";
814 let RenderMethod = "addImmOperands";
818 def EndpgmMatchClass : AsmOperandClass {
819 let Name = "EndpgmImm";
820 let PredicateMethod = "isEndpgm";
821 let ParserMethod = "parseEndpgmOp";
822 let RenderMethod = "addImmOperands";
826 def ExpTgtMatchClass : AsmOperandClass {
828 let PredicateMethod = "isExpTgt";
829 let ParserMethod = "parseExpTgt";
830 let RenderMethod = "printExpTgt";
833 def SendMsgImm : Operand<i32> {
834 let PrintMethod = "printSendMsg";
835 let ParserMatchClass = SendMsgMatchClass;
838 def SwizzleImm : Operand<i16> {
839 let PrintMethod = "printSwizzle";
840 let ParserMatchClass = SwizzleMatchClass;
843 def EndpgmImm : Operand<i16> {
844 let PrintMethod = "printEndpgm";
845 let ParserMatchClass = EndpgmMatchClass;
848 def SWaitMatchClass : AsmOperandClass {
849 let Name = "SWaitCnt";
850 let RenderMethod = "addImmOperands";
851 let ParserMethod = "parseSWaitCntOps";
854 def VReg32OrOffClass : AsmOperandClass {
855 let Name = "VReg32OrOff";
856 let ParserMethod = "parseVReg32OrOff";
859 def WAIT_FLAG : Operand <i32> {
860 let ParserMatchClass = SWaitMatchClass;
861 let PrintMethod = "printWaitFlag";
862 let OperandType = "OPERAND_IMMEDIATE";
865 include "SIInstrFormats.td"
866 include "VIInstrFormats.td"
868 def BoolReg : AsmOperandClass {
869 let Name = "BoolReg";
870 let ParserMethod = "parseBoolReg";
871 let RenderMethod = "addRegOperands";
874 class BoolRC : RegisterOperand<SReg_1> {
875 let ParserMatchClass = BoolReg;
876 let DecoderMethod = "decodeBoolReg";
879 def SSrc_i1 : RegisterOperand<SReg_1_XEXEC> {
880 let ParserMatchClass = BoolReg;
881 let DecoderMethod = "decodeBoolReg";
884 def VOPDstS64orS32 : BoolRC {
885 let PrintMethod = "printVOPDst";
888 // SCSrc_i1 is the operand for pseudo instructions only.
889 // Boolean immeadiates shall not be exposed to codegen instructions.
890 def SCSrc_i1 : RegisterOperand<SReg_1_XEXEC> {
891 let OperandNamespace = "AMDGPU";
892 let OperandType = "OPERAND_REG_IMM_INT32";
893 let ParserMatchClass = BoolReg;
894 let DecoderMethod = "decodeBoolReg";
897 // ===----------------------------------------------------------------------===//
898 // ExpSrc* Special cases for exp src operands which are printed as
899 // "off" depending on en operand.
900 // ===----------------------------------------------------------------------===//
902 def ExpSrc0 : RegisterOperand<VGPR_32> {
903 let PrintMethod = "printExpSrc0";
904 let ParserMatchClass = VReg32OrOffClass;
907 def ExpSrc1 : RegisterOperand<VGPR_32> {
908 let PrintMethod = "printExpSrc1";
909 let ParserMatchClass = VReg32OrOffClass;
912 def ExpSrc2 : RegisterOperand<VGPR_32> {
913 let PrintMethod = "printExpSrc2";
914 let ParserMatchClass = VReg32OrOffClass;
917 def ExpSrc3 : RegisterOperand<VGPR_32> {
918 let PrintMethod = "printExpSrc3";
919 let ParserMatchClass = VReg32OrOffClass;
922 class SDWASrc<ValueType vt> : RegisterOperand<VS_32> {
923 let OperandNamespace = "AMDGPU";
924 string Type = !if(isFloatType<vt>.ret, "FP", "INT");
925 let OperandType = "OPERAND_REG_INLINE_C_"#Type#vt.Size;
926 let DecoderMethod = "decodeSDWASrc"#vt.Size;
927 let EncoderMethod = "getSDWASrcEncoding";
930 def SDWASrc_i32 : SDWASrc<i32>;
931 def SDWASrc_i16 : SDWASrc<i16>;
932 def SDWASrc_f32 : SDWASrc<f32>;
933 def SDWASrc_f16 : SDWASrc<f16>;
935 def SDWAVopcDst : BoolRC {
936 let OperandNamespace = "AMDGPU";
937 let OperandType = "OPERAND_SDWA_VOPC_DST";
938 let EncoderMethod = "getSDWAVopcDstEncoding";
939 let DecoderMethod = "decodeSDWAVopcDst";
940 let PrintMethod = "printVOPDst";
943 class NamedMatchClass<string CName, bit Optional = 1> : AsmOperandClass {
944 let Name = "Imm"#CName;
945 let PredicateMethod = "is"#CName;
946 let ParserMethod = !if(Optional, "parseOptionalOperand", "parse"#CName);
947 let RenderMethod = "addImmOperands";
948 let IsOptional = Optional;
949 let DefaultMethod = !if(Optional, "default"#CName, ?);
952 class NamedOperandBit<string Name, AsmOperandClass MatchClass> : Operand<i1> {
953 let PrintMethod = "print"#Name;
954 let ParserMatchClass = MatchClass;
957 class NamedOperandU8<string Name, AsmOperandClass MatchClass> : Operand<i8> {
958 let PrintMethod = "print"#Name;
959 let ParserMatchClass = MatchClass;
962 class NamedOperandU16<string Name, AsmOperandClass MatchClass> : Operand<i16> {
963 let PrintMethod = "print"#Name;
964 let ParserMatchClass = MatchClass;
967 class NamedOperandU32<string Name, AsmOperandClass MatchClass> : Operand<i32> {
968 let PrintMethod = "print"#Name;
969 let ParserMatchClass = MatchClass;
972 class NamedOperandU32Default0<string Name, AsmOperandClass MatchClass> :
973 OperandWithDefaultOps<i32, (ops (i32 0))> {
974 let PrintMethod = "print"#Name;
975 let ParserMatchClass = MatchClass;
978 let OperandType = "OPERAND_IMMEDIATE" in {
980 def offen : NamedOperandBit<"Offen", NamedMatchClass<"Offen">>;
981 def idxen : NamedOperandBit<"Idxen", NamedMatchClass<"Idxen">>;
982 def addr64 : NamedOperandBit<"Addr64", NamedMatchClass<"Addr64">>;
984 def flat_offset : NamedOperandU16<"FlatOffset", NamedMatchClass<"FlatOffset">>;
985 def offset : NamedOperandU16<"Offset", NamedMatchClass<"Offset">>;
986 def offset0 : NamedOperandU8<"Offset0", NamedMatchClass<"Offset0">>;
987 def offset1 : NamedOperandU8<"Offset1", NamedMatchClass<"Offset1">>;
989 def gds : NamedOperandBit<"GDS", NamedMatchClass<"GDS">>;
991 def omod : NamedOperandU32<"OModSI", NamedMatchClass<"OModSI">>;
992 def clampmod : NamedOperandBit<"ClampSI", NamedMatchClass<"ClampSI">>;
993 def highmod : NamedOperandBit<"High", NamedMatchClass<"High">>;
995 def DLC : NamedOperandBit<"DLC", NamedMatchClass<"DLC">>;
996 def GLC : NamedOperandBit<"GLC", NamedMatchClass<"GLC">>;
997 def SLC : NamedOperandBit<"SLC", NamedMatchClass<"SLC">>;
998 def TFE : NamedOperandBit<"TFE", NamedMatchClass<"TFE">>;
999 def UNorm : NamedOperandBit<"UNorm", NamedMatchClass<"UNorm">>;
1000 def DA : NamedOperandBit<"DA", NamedMatchClass<"DA">>;
1001 def R128A16 : NamedOperandBit<"R128A16", NamedMatchClass<"R128A16">>;
1002 def D16 : NamedOperandBit<"D16", NamedMatchClass<"D16">>;
1003 def LWE : NamedOperandBit<"LWE", NamedMatchClass<"LWE">>;
1004 def exp_compr : NamedOperandBit<"ExpCompr", NamedMatchClass<"ExpCompr">>;
1005 def exp_vm : NamedOperandBit<"ExpVM", NamedMatchClass<"ExpVM">>;
1007 def FORMAT : NamedOperandU8<"FORMAT", NamedMatchClass<"FORMAT">>;
1009 def DMask : NamedOperandU16<"DMask", NamedMatchClass<"DMask">>;
1010 def Dim : NamedOperandU8<"Dim", NamedMatchClass<"Dim", 0>>;
1012 def dpp8 : NamedOperandU32<"DPP8", NamedMatchClass<"DPP8", 0>>;
1014 def dpp_ctrl : NamedOperandU32<"DPPCtrl", NamedMatchClass<"DPPCtrl", 0>>;
1015 def row_mask : NamedOperandU32<"RowMask", NamedMatchClass<"RowMask">>;
1016 def bank_mask : NamedOperandU32<"BankMask", NamedMatchClass<"BankMask">>;
1017 def bound_ctrl : NamedOperandBit<"BoundCtrl", NamedMatchClass<"BoundCtrl">>;
1018 def FI : NamedOperandU32<"FI", NamedMatchClass<"FI">>;
1020 def dst_sel : NamedOperandU32<"SDWADstSel", NamedMatchClass<"SDWADstSel">>;
1021 def src0_sel : NamedOperandU32<"SDWASrc0Sel", NamedMatchClass<"SDWASrc0Sel">>;
1022 def src1_sel : NamedOperandU32<"SDWASrc1Sel", NamedMatchClass<"SDWASrc1Sel">>;
1023 def dst_unused : NamedOperandU32<"SDWADstUnused", NamedMatchClass<"SDWADstUnused">>;
1025 def op_sel : NamedOperandU32Default0<"OpSel", NamedMatchClass<"OpSel">>;
1026 def op_sel_hi : NamedOperandU32Default0<"OpSelHi", NamedMatchClass<"OpSelHi">>;
1027 def neg_lo : NamedOperandU32Default0<"NegLo", NamedMatchClass<"NegLo">>;
1028 def neg_hi : NamedOperandU32Default0<"NegHi", NamedMatchClass<"NegHi">>;
1030 def blgp : NamedOperandU32<"BLGP", NamedMatchClass<"BLGP">>;
1031 def cbsz : NamedOperandU32<"CBSZ", NamedMatchClass<"CBSZ">>;
1032 def abid : NamedOperandU32<"ABID", NamedMatchClass<"ABID">>;
1034 def hwreg : NamedOperandU16<"Hwreg", NamedMatchClass<"Hwreg", 0>>;
1036 def exp_tgt : NamedOperandU8<"ExpTgt", NamedMatchClass<"ExpTgt", 0>> {
1040 } // End OperandType = "OPERAND_IMMEDIATE"
1042 class KImmMatchClass<int size> : AsmOperandClass {
1043 let Name = "KImmFP"#size;
1044 let PredicateMethod = "isKImmFP"#size;
1045 let ParserMethod = "parseImm";
1046 let RenderMethod = "addKImmFP"#size#"Operands";
1049 class kimmOperand<ValueType vt> : Operand<vt> {
1050 let OperandNamespace = "AMDGPU";
1051 let OperandType = "OPERAND_KIMM"#vt.Size;
1052 let PrintMethod = "printU"#vt.Size#"ImmOperand";
1053 let ParserMatchClass = !cast<AsmOperandClass>("KImmFP"#vt.Size#"MatchClass");
1056 // 32-bit VALU immediate operand that uses the constant bus.
1057 def KImmFP32MatchClass : KImmMatchClass<32>;
1058 def f32kimm : kimmOperand<i32>;
1060 // 32-bit VALU immediate operand with a 16-bit value that uses the
1062 def KImmFP16MatchClass : KImmMatchClass<16>;
1063 def f16kimm : kimmOperand<i16>;
1065 class FPInputModsMatchClass <int opSize> : AsmOperandClass {
1066 let Name = "RegOrImmWithFP"#opSize#"InputMods";
1067 let ParserMethod = "parseRegOrImmWithFPInputMods";
1068 let PredicateMethod = "isRegOrImmWithFP"#opSize#"InputMods";
1071 def FP16InputModsMatchClass : FPInputModsMatchClass<16>;
1072 def FP32InputModsMatchClass : FPInputModsMatchClass<32>;
1073 def FP64InputModsMatchClass : FPInputModsMatchClass<64>;
1075 class InputMods <AsmOperandClass matchClass> : Operand <i32> {
1076 let OperandNamespace = "AMDGPU";
1077 let OperandType = "OPERAND_INPUT_MODS";
1078 let ParserMatchClass = matchClass;
1081 class FPInputMods <FPInputModsMatchClass matchClass> : InputMods <matchClass> {
1082 let PrintMethod = "printOperandAndFPInputMods";
1085 def FP16InputMods : FPInputMods<FP16InputModsMatchClass>;
1086 def FP32InputMods : FPInputMods<FP32InputModsMatchClass>;
1087 def FP64InputMods : FPInputMods<FP64InputModsMatchClass>;
1089 class IntInputModsMatchClass <int opSize> : AsmOperandClass {
1090 let Name = "RegOrImmWithInt"#opSize#"InputMods";
1091 let ParserMethod = "parseRegOrImmWithIntInputMods";
1092 let PredicateMethod = "isRegOrImmWithInt"#opSize#"InputMods";
1094 def Int32InputModsMatchClass : IntInputModsMatchClass<32>;
1095 def Int64InputModsMatchClass : IntInputModsMatchClass<64>;
1097 class IntInputMods <IntInputModsMatchClass matchClass> : InputMods <matchClass> {
1098 let PrintMethod = "printOperandAndIntInputMods";
1100 def Int32InputMods : IntInputMods<Int32InputModsMatchClass>;
1101 def Int64InputMods : IntInputMods<Int64InputModsMatchClass>;
1103 class OpSelModsMatchClass : AsmOperandClass {
1104 let Name = "OpSelMods";
1105 let ParserMethod = "parseRegOrImm";
1106 let PredicateMethod = "isRegOrImm";
1109 def IntOpSelModsMatchClass : OpSelModsMatchClass;
1110 def IntOpSelMods : InputMods<IntOpSelModsMatchClass>;
1112 class FPSDWAInputModsMatchClass <int opSize> : AsmOperandClass {
1113 let Name = "SDWAWithFP"#opSize#"InputMods";
1114 let ParserMethod = "parseRegOrImmWithFPInputMods";
1115 let PredicateMethod = "isSDWAFP"#opSize#"Operand";
1118 def FP16SDWAInputModsMatchClass : FPSDWAInputModsMatchClass<16>;
1119 def FP32SDWAInputModsMatchClass : FPSDWAInputModsMatchClass<32>;
1121 class FPSDWAInputMods <FPSDWAInputModsMatchClass matchClass> :
1122 InputMods <matchClass> {
1123 let PrintMethod = "printOperandAndFPInputMods";
1126 def FP16SDWAInputMods : FPSDWAInputMods<FP16SDWAInputModsMatchClass>;
1127 def FP32SDWAInputMods : FPSDWAInputMods<FP32SDWAInputModsMatchClass>;
1129 def FPVRegInputModsMatchClass : AsmOperandClass {
1130 let Name = "VRegWithFPInputMods";
1131 let ParserMethod = "parseRegWithFPInputMods";
1132 let PredicateMethod = "isVReg32";
1135 def FPVRegInputMods : InputMods <FPVRegInputModsMatchClass> {
1136 let PrintMethod = "printOperandAndFPInputMods";
1139 class IntSDWAInputModsMatchClass <int opSize> : AsmOperandClass {
1140 let Name = "SDWAWithInt"#opSize#"InputMods";
1141 let ParserMethod = "parseRegOrImmWithIntInputMods";
1142 let PredicateMethod = "isSDWAInt"#opSize#"Operand";
1145 def Int16SDWAInputModsMatchClass : IntSDWAInputModsMatchClass<16>;
1146 def Int32SDWAInputModsMatchClass : IntSDWAInputModsMatchClass<32>;
1148 class IntSDWAInputMods <IntSDWAInputModsMatchClass matchClass> :
1149 InputMods <matchClass> {
1150 let PrintMethod = "printOperandAndIntInputMods";
1153 def Int16SDWAInputMods : IntSDWAInputMods<Int16SDWAInputModsMatchClass>;
1154 def Int32SDWAInputMods : IntSDWAInputMods<Int32SDWAInputModsMatchClass>;
1156 def IntVRegInputModsMatchClass : AsmOperandClass {
1157 let Name = "VRegWithIntInputMods";
1158 let ParserMethod = "parseRegWithIntInputMods";
1159 let PredicateMethod = "isVReg32";
1162 def IntVRegInputMods : InputMods <IntVRegInputModsMatchClass> {
1163 let PrintMethod = "printOperandAndIntInputMods";
1166 class PackedFPInputModsMatchClass <int opSize> : AsmOperandClass {
1167 let Name = "PackedFP"#opSize#"InputMods";
1168 let ParserMethod = "parseRegOrImm";
1169 let PredicateMethod = "isRegOrImm";
1170 // let PredicateMethod = "isPackedFP"#opSize#"InputMods";
1173 class PackedIntInputModsMatchClass <int opSize> : AsmOperandClass {
1174 let Name = "PackedInt"#opSize#"InputMods";
1175 let ParserMethod = "parseRegOrImm";
1176 let PredicateMethod = "isRegOrImm";
1177 // let PredicateMethod = "isPackedInt"#opSize#"InputMods";
1180 def PackedF16InputModsMatchClass : PackedFPInputModsMatchClass<16>;
1181 def PackedI16InputModsMatchClass : PackedIntInputModsMatchClass<16>;
1183 class PackedFPInputMods <PackedFPInputModsMatchClass matchClass> : InputMods <matchClass> {
1184 // let PrintMethod = "printPackedFPInputMods";
1187 class PackedIntInputMods <PackedIntInputModsMatchClass matchClass> : InputMods <matchClass> {
1188 //let PrintMethod = "printPackedIntInputMods";
1191 def PackedF16InputMods : PackedFPInputMods<PackedF16InputModsMatchClass>;
1192 def PackedI16InputMods : PackedIntInputMods<PackedI16InputModsMatchClass>;
1194 //===----------------------------------------------------------------------===//
1196 //===----------------------------------------------------------------------===//
1198 def DS1Addr1Offset : ComplexPattern<i32, 2, "SelectDS1Addr1Offset">;
1199 def DS64Bit4ByteAligned : ComplexPattern<i32, 3, "SelectDS64Bit4ByteAligned">;
1201 def MOVRELOffset : ComplexPattern<i32, 2, "SelectMOVRELOffset">;
1203 def VOP3Mods0 : ComplexPattern<untyped, 4, "SelectVOP3Mods0">;
1204 def VOP3Mods0Clamp : ComplexPattern<untyped, 3, "SelectVOP3Mods0Clamp">;
1205 def VOP3Mods0Clamp0OMod : ComplexPattern<untyped, 4, "SelectVOP3Mods0Clamp0OMod">;
1206 def VOP3Mods : ComplexPattern<untyped, 2, "SelectVOP3Mods">;
1207 def VOP3NoMods : ComplexPattern<untyped, 1, "SelectVOP3NoMods">;
1208 // VOP3Mods, but the input source is known to never be NaN.
1209 def VOP3Mods_nnan : ComplexPattern<fAny, 2, "SelectVOP3Mods_NNaN">;
1210 // VOP3Mods, but only allowed for f32 operands.
1211 def VOP3Mods_f32 : ComplexPattern<fAny, 2, "SelectVOP3Mods_f32">;
1213 def VOP3OMods : ComplexPattern<untyped, 3, "SelectVOP3OMods">;
1215 def VOP3PMods : ComplexPattern<untyped, 2, "SelectVOP3PMods">;
1216 def VOP3PMods0 : ComplexPattern<untyped, 3, "SelectVOP3PMods0">;
1218 def VOP3OpSel : ComplexPattern<untyped, 2, "SelectVOP3OpSel">;
1219 def VOP3OpSel0 : ComplexPattern<untyped, 3, "SelectVOP3OpSel0">;
1221 def VOP3OpSelMods : ComplexPattern<untyped, 2, "SelectVOP3OpSelMods">;
1222 def VOP3OpSelMods0 : ComplexPattern<untyped, 3, "SelectVOP3OpSelMods0">;
1224 def VOP3PMadMixMods : ComplexPattern<untyped, 2, "SelectVOP3PMadMixMods">;
1227 def Hi16Elt : ComplexPattern<untyped, 1, "SelectHi16Elt">;
1229 //===----------------------------------------------------------------------===//
1230 // SI assembler operands
1231 //===----------------------------------------------------------------------===//
1236 int FLAT_SCR = 0x68;
1239 // This should be kept in sync with SISrcMods enum
1263 int LLVM_DEBUG_TRAP = 3;
1279 int FLAT_SCR_LO = 20;
1280 int FLAT_SCR_HI = 21;
1281 int XNACK_MASK = 22;
1282 int POPS_PACKER = 25;
1285 class getHwRegImm<int Reg, int Offset = 0, int Size = 32> {
1287 !or(!shl(Offset, 6),
1288 !shl(!add(Size, -1), 11)));
1291 //===----------------------------------------------------------------------===//
1293 // SI Instruction multiclass helpers.
1295 // Instructions with _32 take 32-bit operands.
1296 // Instructions with _64 take 64-bit operands.
1298 // VOP_* instructions can use either a 32-bit or 64-bit encoding. The 32-bit
1299 // encoding is the standard encoding, but instruction that make use of
1300 // any of the instruction modifiers must use the 64-bit encoding.
1302 // Instructions with _e32 use the 32-bit encoding.
1303 // Instructions with _e64 use the 64-bit encoding.
1305 //===----------------------------------------------------------------------===//
1307 class SIMCInstr <string pseudo, int subtarget> {
1308 string PseudoInstr = pseudo;
1309 int Subtarget = subtarget;
1312 //===----------------------------------------------------------------------===//
1314 //===----------------------------------------------------------------------===//
1316 class EXP_Helper<bit done, SDPatternOperator node = null_frag> : EXPCommon<
1319 ExpSrc0:$src0, ExpSrc1:$src1, ExpSrc2:$src2, ExpSrc3:$src3,
1320 exp_vm:$vm, exp_compr:$compr, i8imm:$en),
1321 "exp$tgt $src0, $src1, $src2, $src3"#!if(done, " done", "")#"$compr$vm",
1322 [(node (i8 timm:$tgt), (i8 timm:$en),
1323 f32:$src0, f32:$src1, f32:$src2, f32:$src3,
1324 (i1 timm:$compr), (i1 timm:$vm))]> {
1325 let AsmMatchConverter = "cvtExp";
1328 // Split EXP instruction into EXP and EXP_DONE so we can set
1329 // mayLoad for done=1.
1330 multiclass EXP_m<bit done, SDPatternOperator node> {
1331 let mayLoad = done, DisableWQM = 1 in {
1332 let isPseudo = 1, isCodeGenOnly = 1 in {
1333 def "" : EXP_Helper<done, node>,
1334 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.NONE>;
1337 let done = done in {
1338 def _si : EXP_Helper<done>,
1339 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.SI>,
1341 let AssemblerPredicates = [isGFX6GFX7];
1342 let DecoderNamespace = "GFX6GFX7";
1343 let DisableDecoder = DisableSIDecoder;
1346 def _vi : EXP_Helper<done>,
1347 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.VI>,
1349 let AssemblerPredicates = [isGFX8GFX9];
1350 let DecoderNamespace = "GFX8";
1351 let DisableDecoder = DisableVIDecoder;
1354 def _gfx10 : EXP_Helper<done>,
1355 SIMCInstr <"exp"#!if(done, "_done", ""), SIEncodingFamily.GFX10>,
1357 let AssemblerPredicates = [isGFX10Plus];
1358 let DecoderNamespace = "GFX10";
1359 let DisableDecoder = DisableSIDecoder;
1365 //===----------------------------------------------------------------------===//
1366 // Vector ALU classes
1367 //===----------------------------------------------------------------------===//
1369 class getNumSrcArgs<ValueType Src0, ValueType Src1, ValueType Src2> {
1371 !if (!eq(Src0.Value, untyped.Value), 0,
1372 !if (!eq(Src1.Value, untyped.Value), 1, // VOP1
1373 !if (!eq(Src2.Value, untyped.Value), 2, // VOP2
1377 // Returns the register class to use for the destination of VOP[123C]
1378 // instructions for the given VT.
1379 class getVALUDstForVT<ValueType VT> {
1380 RegisterOperand ret = !if(!eq(VT.Size, 32), VOPDstOperand<VGPR_32>,
1381 !if(!eq(VT.Size, 128), VOPDstOperand<VReg_128>,
1382 !if(!eq(VT.Size, 64), VOPDstOperand<VReg_64>,
1383 !if(!eq(VT.Size, 16), VOPDstOperand<VGPR_32>,
1384 VOPDstS64orS32)))); // else VT == i1
1387 // Returns true if VT is floating point.
1388 class getIsFP<ValueType VT> {
1389 bit ret = !if(!eq(VT.Value, f16.Value), 1,
1390 !if(!eq(VT.Value, v2f16.Value), 1,
1391 !if(!eq(VT.Value, v4f16.Value), 1,
1392 !if(!eq(VT.Value, f32.Value), 1,
1393 !if(!eq(VT.Value, v2f32.Value), 1,
1394 !if(!eq(VT.Value, f64.Value), 1,
1395 !if(!eq(VT.Value, v2f64.Value), 1,
1399 // Returns the register class to use for the destination of VOP[12C]
1400 // instructions with SDWA extension
1401 class getSDWADstForVT<ValueType VT> {
1402 RegisterOperand ret = !if(!eq(VT.Size, 1),
1403 SDWAVopcDst, // VOPC
1404 VOPDstOperand<VGPR_32>); // VOP1/2 32-bit dst
1407 // Returns the register class to use for source 0 of VOP[12C]
1408 // instructions for the given VT.
1409 class getVOPSrc0ForVT<ValueType VT> {
1410 bit isFP = getIsFP<VT>.ret;
1412 RegisterOperand ret =
1414 !if(!eq(VT.Size, 64),
1416 !if(!eq(VT.Value, f16.Value),
1418 !if(!eq(VT.Value, v2f16.Value),
1420 !if(!eq(VT.Value, v4f16.Value),
1427 !if(!eq(VT.Size, 64),
1429 !if(!eq(VT.Value, i16.Value),
1431 !if(!eq(VT.Value, v2i16.Value),
1440 // Returns the vreg register class to use for source operand given VT
1441 class getVregSrcForVT<ValueType VT> {
1442 RegisterClass ret = !if(!eq(VT.Size, 128), VReg_128,
1443 !if(!eq(VT.Size, 96), VReg_96,
1444 !if(!eq(VT.Size, 64), VReg_64,
1445 !if(!eq(VT.Size, 48), VReg_64,
1449 class getSDWASrcForVT <ValueType VT> {
1450 bit isFP = getIsFP<VT>.ret;
1451 RegisterOperand retFlt = !if(!eq(VT.Size, 16), SDWASrc_f16, SDWASrc_f32);
1452 RegisterOperand retInt = !if(!eq(VT.Size, 16), SDWASrc_i16, SDWASrc_i32);
1453 RegisterOperand ret = !if(isFP, retFlt, retInt);
1456 // Returns the register class to use for sources of VOP3 instructions for the
1458 class getVOP3SrcForVT<ValueType VT> {
1459 bit isFP = getIsFP<VT>.ret;
1460 RegisterOperand ret =
1461 !if(!eq(VT.Size, 128),
1463 !if(!eq(VT.Size, 64),
1467 !if(!eq(VT.Value, i1.Value),
1470 !if(!eq(VT.Value, f16.Value),
1472 !if(!eq(VT.Value, v2f16.Value),
1474 !if(!eq(VT.Value, v4f16.Value),
1480 !if(!eq(VT.Value, i16.Value),
1482 !if(!eq(VT.Value, v2i16.Value),
1493 // Float or packed int
1494 class isModifierType<ValueType SrcVT> {
1496 !if(!eq(SrcVT.Value, f16.Value), 1,
1497 !if(!eq(SrcVT.Value, f32.Value), 1,
1498 !if(!eq(SrcVT.Value, f64.Value), 1,
1499 !if(!eq(SrcVT.Value, v2f16.Value), 1,
1500 !if(!eq(SrcVT.Value, v2i16.Value), 1,
1504 // Return type of input modifiers operand for specified input operand
1505 class getSrcMod <ValueType VT, bit EnableF32SrcMods> {
1506 bit isFP = getIsFP<VT>.ret;
1507 bit isPacked = isPackedType<VT>.ret;
1508 Operand ret = !if(!eq(VT.Size, 64),
1509 !if(isFP, FP64InputMods, Int64InputMods),
1511 !if(!eq(VT.Value, f16.Value),
1515 !if(EnableF32SrcMods, FP32InputMods, Int32InputMods))
1519 class getOpSelMod <ValueType VT> {
1520 Operand ret = !if(!eq(VT.Value, f16.Value), FP16InputMods, IntOpSelMods);
1523 // Return type of input modifiers operand specified input operand for DPP
1524 class getSrcModExt <ValueType VT> {
1525 bit isFP = getIsFP<VT>.ret;
1526 Operand ret = !if(isFP, FPVRegInputMods, IntVRegInputMods);
1529 // Return type of input modifiers operand specified input operand for SDWA
1530 class getSrcModSDWA <ValueType VT> {
1531 Operand ret = !if(!eq(VT.Value, f16.Value), FP16SDWAInputMods,
1532 !if(!eq(VT.Value, f32.Value), FP32SDWAInputMods,
1533 !if(!eq(VT.Value, i16.Value), Int16SDWAInputMods,
1534 Int32SDWAInputMods)));
1537 // Returns the input arguments for VOP[12C] instructions for the given SrcVT.
1538 class getIns32 <RegisterOperand Src0RC, RegisterClass Src1RC, int NumSrcArgs> {
1539 dag ret = !if(!eq(NumSrcArgs, 1), (ins Src0RC:$src0), // VOP1
1540 !if(!eq(NumSrcArgs, 2), (ins Src0RC:$src0, Src1RC:$src1), // VOP2
1544 // Returns the input arguments for VOP3 instructions for the given SrcVT.
1545 class getIns64 <RegisterOperand Src0RC, RegisterOperand Src1RC,
1546 RegisterOperand Src2RC, int NumSrcArgs,
1547 bit HasIntClamp, bit HasModifiers, bit HasSrc2Mods, bit HasOMod,
1548 Operand Src0Mod, Operand Src1Mod, Operand Src2Mod> {
1551 !if (!eq(NumSrcArgs, 0),
1552 // VOP1 without input operands (V_NOP, V_CLREXCP)
1555 !if (!eq(NumSrcArgs, 1),
1556 !if (!eq(HasModifiers, 1),
1557 // VOP1 with modifiers
1558 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1559 clampmod:$clamp, omod:$omod)
1561 // VOP1 without modifiers
1562 !if (!eq(HasIntClamp, 1),
1563 (ins Src0RC:$src0, clampmod:$clamp),
1566 !if (!eq(NumSrcArgs, 2),
1567 !if (!eq(HasModifiers, 1),
1568 // VOP 2 with modifiers
1569 !if( !eq(HasOMod, 1),
1570 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1571 Src1Mod:$src1_modifiers, Src1RC:$src1,
1572 clampmod:$clamp, omod:$omod),
1573 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1574 Src1Mod:$src1_modifiers, Src1RC:$src1,
1577 // VOP2 without modifiers
1578 !if (!eq(HasIntClamp, 1),
1579 (ins Src0RC:$src0, Src1RC:$src1, clampmod:$clamp),
1580 (ins Src0RC:$src0, Src1RC:$src1))
1583 /* NumSrcArgs == 3 */,
1584 !if (!eq(HasModifiers, 1),
1585 !if (!eq(HasSrc2Mods, 1),
1586 // VOP3 with modifiers
1587 !if (!eq(HasOMod, 1),
1588 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1589 Src1Mod:$src1_modifiers, Src1RC:$src1,
1590 Src2Mod:$src2_modifiers, Src2RC:$src2,
1591 clampmod:$clamp, omod:$omod),
1592 !if (!eq(HasIntClamp, 1),
1593 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1594 Src1Mod:$src1_modifiers, Src1RC:$src1,
1595 Src2Mod:$src2_modifiers, Src2RC:$src2,
1597 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1598 Src1Mod:$src1_modifiers, Src1RC:$src1,
1599 Src2Mod:$src2_modifiers, Src2RC:$src2))),
1600 // VOP3 with modifiers except src2
1601 !if (!eq(HasOMod, 1),
1602 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1603 Src1Mod:$src1_modifiers, Src1RC:$src1,
1604 Src2RC:$src2, clampmod:$clamp, omod:$omod),
1605 !if (!eq(HasIntClamp, 1),
1606 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1607 Src1Mod:$src1_modifiers, Src1RC:$src1,
1608 Src2RC:$src2, clampmod:$clamp),
1609 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1610 Src1Mod:$src1_modifiers, Src1RC:$src1,
1613 // VOP3 without modifiers
1614 !if (!eq(HasIntClamp, 1),
1615 (ins Src0RC:$src0, Src1RC:$src1, Src2RC:$src2, clampmod:$clamp),
1616 (ins Src0RC:$src0, Src1RC:$src1, Src2RC:$src2))
1620 /// XXX - src1 may only allow VGPRs?
1622 // The modifiers (except clamp) are dummy operands for the benefit of
1623 // printing and parsing. They defer their values to looking at the
1624 // srcN_modifiers for what to print.
1625 class getInsVOP3P <RegisterOperand Src0RC, RegisterOperand Src1RC,
1626 RegisterOperand Src2RC, int NumSrcArgs,
1628 Operand Src0Mod, Operand Src1Mod, Operand Src2Mod> {
1629 dag ret = !if (!eq(NumSrcArgs, 2),
1631 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1632 Src1Mod:$src1_modifiers, Src1RC:$src1,
1634 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1635 neg_lo:$neg_lo, neg_hi:$neg_hi),
1636 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1637 Src1Mod:$src1_modifiers, Src1RC:$src1,
1638 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1639 neg_lo:$neg_lo, neg_hi:$neg_hi)),
1640 // else NumSrcArgs == 3
1642 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1643 Src1Mod:$src1_modifiers, Src1RC:$src1,
1644 Src2Mod:$src2_modifiers, Src2RC:$src2,
1646 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1647 neg_lo:$neg_lo, neg_hi:$neg_hi),
1648 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1649 Src1Mod:$src1_modifiers, Src1RC:$src1,
1650 Src2Mod:$src2_modifiers, Src2RC:$src2,
1651 op_sel:$op_sel, op_sel_hi:$op_sel_hi,
1652 neg_lo:$neg_lo, neg_hi:$neg_hi))
1656 class getInsVOP3OpSel <RegisterOperand Src0RC,
1657 RegisterOperand Src1RC,
1658 RegisterOperand Src2RC,
1664 dag ret = !if (!eq(NumSrcArgs, 2),
1666 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1667 Src1Mod:$src1_modifiers, Src1RC:$src1,
1670 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1671 Src1Mod:$src1_modifiers, Src1RC:$src1,
1673 // else NumSrcArgs == 3
1675 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1676 Src1Mod:$src1_modifiers, Src1RC:$src1,
1677 Src2Mod:$src2_modifiers, Src2RC:$src2,
1680 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1681 Src1Mod:$src1_modifiers, Src1RC:$src1,
1682 Src2Mod:$src2_modifiers, Src2RC:$src2,
1687 class getInsDPP <RegisterOperand DstRC, RegisterClass Src0RC, RegisterClass Src1RC,
1688 int NumSrcArgs, bit HasModifiers,
1689 Operand Src0Mod, Operand Src1Mod> {
1691 dag ret = !if (!eq(NumSrcArgs, 0),
1692 // VOP1 without input operands (V_NOP)
1693 (ins dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1694 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl),
1695 !if (!eq(NumSrcArgs, 1),
1696 !if (!eq(HasModifiers, 1),
1697 // VOP1_DPP with modifiers
1698 (ins DstRC:$old, Src0Mod:$src0_modifiers,
1699 Src0RC:$src0, dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1700 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl)
1702 // VOP1_DPP without modifiers
1703 (ins DstRC:$old, Src0RC:$src0,
1704 dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1705 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl)
1707 /* NumSrcArgs == 2 */,
1708 !if (!eq(HasModifiers, 1),
1709 // VOP2_DPP with modifiers
1711 Src0Mod:$src0_modifiers, Src0RC:$src0,
1712 Src1Mod:$src1_modifiers, Src1RC:$src1,
1713 dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
1714 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl)
1716 // VOP2_DPP without modifiers
1718 Src0RC:$src0, Src1RC:$src1, dpp_ctrl:$dpp_ctrl,
1719 row_mask:$row_mask, bank_mask:$bank_mask,
1720 bound_ctrl:$bound_ctrl)
1724 class getInsDPP16 <RegisterOperand DstRC, RegisterClass Src0RC, RegisterClass Src1RC,
1725 int NumSrcArgs, bit HasModifiers,
1726 Operand Src0Mod, Operand Src1Mod> {
1727 dag ret = !con(getInsDPP<DstRC, Src0RC, Src1RC, NumSrcArgs,
1728 HasModifiers, Src0Mod, Src1Mod>.ret,
1732 class getInsDPP8 <RegisterOperand DstRC, RegisterClass Src0RC, RegisterClass Src1RC,
1733 int NumSrcArgs, bit HasModifiers,
1734 Operand Src0Mod, Operand Src1Mod> {
1735 dag ret = !if (!eq(NumSrcArgs, 0),
1736 // VOP1 without input operands (V_NOP)
1737 (ins dpp8:$dpp8, FI:$fi),
1738 !if (!eq(NumSrcArgs, 1),
1739 !if (!eq(HasModifiers, 1),
1740 // VOP1_DPP with modifiers
1741 (ins DstRC:$old, Src0Mod:$src0_modifiers,
1742 Src0RC:$src0, dpp8:$dpp8, FI:$fi)
1744 // VOP1_DPP without modifiers
1745 (ins DstRC:$old, Src0RC:$src0, dpp8:$dpp8, FI:$fi)
1747 /* NumSrcArgs == 2 */,
1748 !if (!eq(HasModifiers, 1),
1749 // VOP2_DPP with modifiers
1751 Src0Mod:$src0_modifiers, Src0RC:$src0,
1752 Src1Mod:$src1_modifiers, Src1RC:$src1,
1755 // VOP2_DPP without modifiers
1757 Src0RC:$src0, Src1RC:$src1, dpp8:$dpp8, FI:$fi)
1763 class getInsSDWA <RegisterOperand Src0RC, RegisterOperand Src1RC, int NumSrcArgs,
1764 bit HasSDWAOMod, Operand Src0Mod, Operand Src1Mod,
1767 dag ret = !if(!eq(NumSrcArgs, 0),
1768 // VOP1 without input operands (V_NOP)
1770 !if(!eq(NumSrcArgs, 1),
1772 !if(!eq(HasSDWAOMod, 0),
1773 // VOP1_SDWA without omod
1774 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1776 dst_sel:$dst_sel, dst_unused:$dst_unused,
1777 src0_sel:$src0_sel),
1778 // VOP1_SDWA with omod
1779 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1780 clampmod:$clamp, omod:$omod,
1781 dst_sel:$dst_sel, dst_unused:$dst_unused,
1782 src0_sel:$src0_sel)),
1783 !if(!eq(NumSrcArgs, 2),
1784 !if(!eq(DstVT.Size, 1),
1786 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1787 Src1Mod:$src1_modifiers, Src1RC:$src1,
1788 clampmod:$clamp, src0_sel:$src0_sel, src1_sel:$src1_sel),
1790 !if(!eq(HasSDWAOMod, 0),
1791 // VOP2_SDWA without omod
1792 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1793 Src1Mod:$src1_modifiers, Src1RC:$src1,
1795 dst_sel:$dst_sel, dst_unused:$dst_unused,
1796 src0_sel:$src0_sel, src1_sel:$src1_sel),
1797 // VOP2_SDWA with omod
1798 (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
1799 Src1Mod:$src1_modifiers, Src1RC:$src1,
1800 clampmod:$clamp, omod:$omod,
1801 dst_sel:$dst_sel, dst_unused:$dst_unused,
1802 src0_sel:$src0_sel, src1_sel:$src1_sel))),
1803 (ins)/* endif */)));
1806 // Outs for DPP and SDWA
1807 class getOutsExt <bit HasDst, ValueType DstVT, RegisterOperand DstRCExt> {
1808 dag ret = !if(HasDst,
1809 !if(!eq(DstVT.Size, 1),
1810 (outs), // no dst for VOPC, we use "vcc"-token as dst in SDWA VOPC instructions
1811 (outs DstRCExt:$vdst)),
1816 class getOutsSDWA <bit HasDst, ValueType DstVT, RegisterOperand DstRCSDWA> {
1817 dag ret = !if(HasDst,
1818 !if(!eq(DstVT.Size, 1),
1819 (outs DstRCSDWA:$sdst),
1820 (outs DstRCSDWA:$vdst)),
1824 // Returns the assembly string for the inputs and outputs of a VOP[12C]
1825 // instruction. This does not add the _e32 suffix, so it can be reused
1827 class getAsm32 <bit HasDst, int NumSrcArgs, ValueType DstVT = i32> {
1828 string dst = !if(!eq(DstVT.Size, 1), "$sdst", "$vdst"); // use $sdst for VOPC
1829 string src0 = ", $src0";
1830 string src1 = ", $src1";
1831 string src2 = ", $src2";
1832 string ret = !if(HasDst, dst, "") #
1833 !if(!eq(NumSrcArgs, 1), src0, "") #
1834 !if(!eq(NumSrcArgs, 2), src0#src1, "") #
1835 !if(!eq(NumSrcArgs, 3), src0#src1#src2, "");
1838 // Returns the assembly string for the inputs and outputs of a VOP3
1840 class getAsm64 <bit HasDst, int NumSrcArgs, bit HasIntClamp, bit HasModifiers,
1841 bit HasOMod, ValueType DstVT = i32> {
1842 string dst = !if(!eq(DstVT.Size, 1), "$sdst", "$vdst"); // use $sdst for VOPC
1843 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1844 string src1 = !if(!eq(NumSrcArgs, 1), "",
1845 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1846 " $src1_modifiers,"));
1847 string src2 = !if(!eq(NumSrcArgs, 3), " $src2_modifiers", "");
1848 string iclamp = !if(HasIntClamp, "$clamp", "");
1850 !if(!eq(HasModifiers, 0),
1851 getAsm32<HasDst, NumSrcArgs, DstVT>.ret # iclamp,
1852 dst#", "#src0#src1#src2#"$clamp"#!if(HasOMod, "$omod", ""));
1855 // Returns the assembly string for the inputs and outputs of a VOP3P
1857 class getAsmVOP3P <bit HasDst, int NumSrcArgs, bit HasModifiers,
1858 bit HasClamp, ValueType DstVT = i32> {
1859 string dst = " $vdst";
1860 string src0 = !if(!eq(NumSrcArgs, 1), "$src0", "$src0,");
1861 string src1 = !if(!eq(NumSrcArgs, 1), "",
1862 !if(!eq(NumSrcArgs, 2), " $src1",
1864 string src2 = !if(!eq(NumSrcArgs, 3), " $src2", "");
1866 string mods = !if(HasModifiers, "$neg_lo$neg_hi", "");
1867 string clamp = !if(HasClamp, "$clamp", "");
1869 // Each modifier is printed as an array of bits for each operand, so
1870 // all operands are printed as part of src0_modifiers.
1871 string ret = dst#", "#src0#src1#src2#"$op_sel$op_sel_hi"#mods#clamp;
1874 class getAsmVOP3OpSel <int NumSrcArgs,
1879 string dst = " $vdst";
1881 string isrc0 = !if(!eq(NumSrcArgs, 1), "$src0", "$src0,");
1882 string isrc1 = !if(!eq(NumSrcArgs, 1), "",
1883 !if(!eq(NumSrcArgs, 2), " $src1",
1885 string isrc2 = !if(!eq(NumSrcArgs, 3), " $src2", "");
1887 string fsrc0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1888 string fsrc1 = !if(!eq(NumSrcArgs, 1), "",
1889 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1890 " $src1_modifiers,"));
1891 string fsrc2 = !if(!eq(NumSrcArgs, 3), " $src2_modifiers", "");
1893 string src0 = !if(Src0HasMods, fsrc0, isrc0);
1894 string src1 = !if(Src1HasMods, fsrc1, isrc1);
1895 string src2 = !if(Src2HasMods, fsrc2, isrc2);
1897 string clamp = !if(HasClamp, "$clamp", "");
1899 string ret = dst#", "#src0#src1#src2#"$op_sel"#clamp;
1902 class getAsmDPP <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> {
1903 string dst = !if(HasDst,
1904 !if(!eq(DstVT.Size, 1),
1907 ""); // use $sdst for VOPC
1908 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1909 string src1 = !if(!eq(NumSrcArgs, 1), "",
1910 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1911 " $src1_modifiers,"));
1912 string args = !if(!eq(HasModifiers, 0),
1913 getAsm32<0, NumSrcArgs, DstVT>.ret,
1915 string ret = dst#args#" $dpp_ctrl$row_mask$bank_mask$bound_ctrl";
1918 class getAsmDPP16 <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> {
1919 string ret = getAsmDPP<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret#"$fi";
1922 class getAsmDPP8 <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> {
1923 string dst = !if(HasDst,
1924 !if(!eq(DstVT.Size, 1),
1927 ""); // use $sdst for VOPC
1928 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,");
1929 string src1 = !if(!eq(NumSrcArgs, 1), "",
1930 !if(!eq(NumSrcArgs, 2), " $src1_modifiers",
1931 " $src1_modifiers,"));
1932 string args = !if(!eq(HasModifiers, 0),
1933 getAsm32<0, NumSrcArgs, DstVT>.ret,
1935 string ret = dst#args#"$dpp8$fi";
1938 class getAsmSDWA <bit HasDst, int NumSrcArgs, ValueType DstVT = i32> {
1939 string dst = !if(HasDst,
1940 !if(!eq(DstVT.Size, 1),
1941 " vcc", // use vcc token as dst for VOPC instructioins
1944 string src0 = "$src0_modifiers";
1945 string src1 = "$src1_modifiers";
1946 string args = !if(!eq(NumSrcArgs, 0),
1948 !if(!eq(NumSrcArgs, 1),
1950 ", "#src0#", "#src1#"$clamp"
1953 string sdwa = !if(!eq(NumSrcArgs, 0),
1955 !if(!eq(NumSrcArgs, 1),
1956 " $dst_sel $dst_unused $src0_sel",
1957 !if(!eq(DstVT.Size, 1),
1958 " $src0_sel $src1_sel", // No dst_sel and dst_unused for VOPC
1959 " $dst_sel $dst_unused $src0_sel $src1_sel"
1963 string ret = dst#args#sdwa;
1966 class getAsmSDWA9 <bit HasDst, bit HasOMod, int NumSrcArgs,
1967 ValueType DstVT = i32> {
1968 string dst = !if(HasDst,
1969 !if(!eq(DstVT.Size, 1),
1973 string src0 = "$src0_modifiers";
1974 string src1 = "$src1_modifiers";
1975 string out_mods = !if(!eq(HasOMod, 0), "$clamp", "$clamp$omod");
1976 string args = !if(!eq(NumSrcArgs, 0), "",
1977 !if(!eq(NumSrcArgs, 1),
1982 string sdwa = !if(!eq(NumSrcArgs, 0), "",
1983 !if(!eq(NumSrcArgs, 1),
1984 out_mods#" $dst_sel $dst_unused $src0_sel",
1985 !if(!eq(DstVT.Size, 1),
1986 " $src0_sel $src1_sel", // No dst_sel, dst_unused and output modifiers for VOPC
1987 out_mods#" $dst_sel $dst_unused $src0_sel $src1_sel"
1991 string ret = dst#args#sdwa;
1995 // Function that checks if instruction supports DPP and SDWA
1996 class getHasExt <int NumSrcArgs, ValueType DstVT = i32, ValueType Src0VT = i32,
1997 ValueType Src1VT = i32> {
1998 bit ret = !if(!eq(NumSrcArgs, 3),
1999 0, // NumSrcArgs == 3 - No DPP or SDWA for VOP3
2000 !if(!eq(DstVT.Size, 64),
2001 0, // 64-bit dst - No DPP or SDWA for 64-bit operands
2002 !if(!eq(Src0VT.Size, 64),
2004 !if(!eq(Src1VT.Size, 64),
2013 class getHasDPP <int NumSrcArgs, ValueType DstVT = i32, ValueType Src0VT = i32,
2014 ValueType Src1VT = i32> {
2015 bit ret = !if(!eq(NumSrcArgs, 0), 0,
2016 getHasExt<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret);
2019 class BitOr<bit a, bit b> {
2020 bit ret = !if(a, 1, !if(b, 1, 0));
2023 class BitAnd<bit a, bit b> {
2024 bit ret = !if(a, !if(b, 1, 0), 0);
2032 class VOPProfile <list<ValueType> _ArgVT, bit _EnableF32SrcMods = 0,
2033 bit _EnableClamp = 0> {
2035 field list<ValueType> ArgVT = _ArgVT;
2036 field bit EnableF32SrcMods = _EnableF32SrcMods;
2037 field bit EnableClamp = _EnableClamp;
2039 field ValueType DstVT = ArgVT[0];
2040 field ValueType Src0VT = ArgVT[1];
2041 field ValueType Src1VT = ArgVT[2];
2042 field ValueType Src2VT = ArgVT[3];
2043 field RegisterOperand DstRC = getVALUDstForVT<DstVT>.ret;
2044 field RegisterOperand DstRCDPP = getVALUDstForVT<DstVT>.ret;
2045 field RegisterOperand DstRCSDWA = getSDWADstForVT<DstVT>.ret;
2046 field RegisterOperand Src0RC32 = getVOPSrc0ForVT<Src0VT>.ret;
2047 field RegisterClass Src1RC32 = getVregSrcForVT<Src1VT>.ret;
2048 field RegisterOperand Src0RC64 = getVOP3SrcForVT<Src0VT>.ret;
2049 field RegisterOperand Src1RC64 = getVOP3SrcForVT<Src1VT>.ret;
2050 field RegisterOperand Src2RC64 = getVOP3SrcForVT<Src2VT>.ret;
2051 field RegisterClass Src0DPP = getVregSrcForVT<Src0VT>.ret;
2052 field RegisterClass Src1DPP = getVregSrcForVT<Src1VT>.ret;
2053 field RegisterOperand Src0SDWA = getSDWASrcForVT<Src0VT>.ret;
2054 field RegisterOperand Src1SDWA = getSDWASrcForVT<Src0VT>.ret;
2055 field Operand Src0Mod = getSrcMod<Src0VT, EnableF32SrcMods>.ret;
2056 field Operand Src1Mod = getSrcMod<Src1VT, EnableF32SrcMods>.ret;
2057 field Operand Src2Mod = getSrcMod<Src2VT, EnableF32SrcMods>.ret;
2058 field Operand Src0ModDPP = getSrcModExt<Src0VT>.ret;
2059 field Operand Src1ModDPP = getSrcModExt<Src1VT>.ret;
2060 field Operand Src0ModSDWA = getSrcModSDWA<Src0VT>.ret;
2061 field Operand Src1ModSDWA = getSrcModSDWA<Src1VT>.ret;
2064 field bit HasDst = !if(!eq(DstVT.Value, untyped.Value), 0, 1);
2065 field bit HasDst32 = HasDst;
2066 field bit EmitDst = HasDst; // force dst encoding, see v_movreld_b32 special case
2067 field int NumSrcArgs = getNumSrcArgs<Src0VT, Src1VT, Src2VT>.ret;
2068 field bit HasSrc0 = !if(!eq(Src0VT.Value, untyped.Value), 0, 1);
2069 field bit HasSrc1 = !if(!eq(Src1VT.Value, untyped.Value), 0, 1);
2070 field bit HasSrc2 = !if(!eq(Src2VT.Value, untyped.Value), 0, 1);
2072 // TODO: Modifiers logic is somewhat adhoc here, to be refined later
2073 // HasModifiers affects the normal and DPP encodings. We take note of EnableF32SrcMods, which
2074 // enables modifiers for i32 type.
2075 field bit HasModifiers = BitOr<isModifierType<Src0VT>.ret, EnableF32SrcMods>.ret;
2077 // HasSrc*FloatMods affects the SDWA encoding. We ignore EnableF32SrcMods.
2078 field bit HasSrc0FloatMods = isFloatType<Src0VT>.ret;
2079 field bit HasSrc1FloatMods = isFloatType<Src1VT>.ret;
2080 field bit HasSrc2FloatMods = isFloatType<Src2VT>.ret;
2082 // HasSrc*IntMods affects the SDWA encoding. We ignore EnableF32SrcMods.
2083 field bit HasSrc0IntMods = isIntType<Src0VT>.ret;
2084 field bit HasSrc1IntMods = isIntType<Src1VT>.ret;
2085 field bit HasSrc2IntMods = isIntType<Src2VT>.ret;
2087 field bit HasSrc0Mods = HasModifiers;
2088 field bit HasSrc1Mods = !if(HasModifiers, BitOr<HasSrc1FloatMods, HasSrc1IntMods>.ret, 0);
2089 field bit HasSrc2Mods = !if(HasModifiers, BitOr<HasSrc2FloatMods, HasSrc2IntMods>.ret, 0);
2091 field bit HasClamp = BitOr<isModifierType<Src0VT>.ret, EnableClamp>.ret;
2092 field bit HasSDWAClamp = EmitDst;
2093 field bit HasFPClamp = BitAnd<isFloatType<DstVT>.ret, HasClamp>.ret;
2094 field bit HasIntClamp = !if(isFloatType<DstVT>.ret, 0, HasClamp);
2095 field bit HasClampLo = HasClamp;
2096 field bit HasClampHi = BitAnd<isPackedType<DstVT>.ret, HasClamp>.ret;
2097 field bit HasHigh = 0;
2099 field bit IsPacked = isPackedType<Src0VT>.ret;
2100 field bit HasOpSel = IsPacked;
2101 field bit HasOMod = !if(HasOpSel, 0, isFloatType<DstVT>.ret);
2102 field bit HasSDWAOMod = isFloatType<DstVT>.ret;
2104 field bit HasExt = getHasExt<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret;
2105 field bit HasExtDPP = getHasDPP<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret;
2106 field bit HasExtSDWA = HasExt;
2107 field bit HasExtSDWA9 = HasExt;
2108 field int NeedPatGen = PatGenMode.NoPattern;
2110 field bit IsMAI = 0;
2112 field Operand Src0PackedMod = !if(HasSrc0FloatMods, PackedF16InputMods, PackedI16InputMods);
2113 field Operand Src1PackedMod = !if(HasSrc1FloatMods, PackedF16InputMods, PackedI16InputMods);
2114 field Operand Src2PackedMod = !if(HasSrc2FloatMods, PackedF16InputMods, PackedI16InputMods);
2116 field dag Outs = !if(HasDst,(outs DstRC:$vdst),(outs));
2118 // VOP3b instructions are a special case with a second explicit
2119 // output. This is manually overridden for them.
2120 field dag Outs32 = Outs;
2121 field dag Outs64 = Outs;
2122 field dag OutsDPP = getOutsExt<HasDst, DstVT, DstRCDPP>.ret;
2123 field dag OutsDPP8 = getOutsExt<HasDst, DstVT, DstRCDPP>.ret;
2124 field dag OutsSDWA = getOutsSDWA<HasDst, DstVT, DstRCSDWA>.ret;
2126 field dag Ins32 = getIns32<Src0RC32, Src1RC32, NumSrcArgs>.ret;
2127 field dag Ins64 = getIns64<Src0RC64, Src1RC64, Src2RC64, NumSrcArgs,
2128 HasIntClamp, HasModifiers, HasSrc2Mods,
2129 HasOMod, Src0Mod, Src1Mod, Src2Mod>.ret;
2130 field dag InsVOP3P = getInsVOP3P<Src0RC64, Src1RC64, Src2RC64,
2131 NumSrcArgs, HasClamp,
2132 Src0PackedMod, Src1PackedMod, Src2PackedMod>.ret;
2133 field dag InsVOP3OpSel = getInsVOP3OpSel<Src0RC64, Src1RC64, Src2RC64,
2136 getOpSelMod<Src0VT>.ret,
2137 getOpSelMod<Src1VT>.ret,
2138 getOpSelMod<Src2VT>.ret>.ret;
2139 field dag InsDPP = !if(HasExtDPP,
2140 getInsDPP<DstRCDPP, Src0DPP, Src1DPP, NumSrcArgs,
2141 HasModifiers, Src0ModDPP, Src1ModDPP>.ret,
2143 field dag InsDPP16 = getInsDPP16<DstRCDPP, Src0DPP, Src1DPP, NumSrcArgs,
2144 HasModifiers, Src0ModDPP, Src1ModDPP>.ret;
2145 field dag InsDPP8 = getInsDPP8<DstRCDPP, Src0DPP, Src1DPP, NumSrcArgs, 0,
2146 Src0ModDPP, Src1ModDPP>.ret;
2147 field dag InsSDWA = getInsSDWA<Src0SDWA, Src1SDWA, NumSrcArgs,
2148 HasSDWAOMod, Src0ModSDWA, Src1ModSDWA,
2152 field string Asm32 = getAsm32<HasDst, NumSrcArgs, DstVT>.ret;
2153 field string Asm64 = getAsm64<HasDst, NumSrcArgs, HasIntClamp, HasModifiers, HasOMod, DstVT>.ret;
2154 field string AsmVOP3P = getAsmVOP3P<HasDst, NumSrcArgs, HasModifiers, HasClamp, DstVT>.ret;
2155 field string AsmVOP3OpSel = getAsmVOP3OpSel<NumSrcArgs,
2159 HasSrc2FloatMods>.ret;
2160 field string AsmDPP = !if(HasExtDPP,
2161 getAsmDPP<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret, "");
2162 field string AsmDPP16 = getAsmDPP16<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret;
2163 field string AsmDPP8 = getAsmDPP8<HasDst, NumSrcArgs, 0, DstVT>.ret;
2164 field string AsmSDWA = getAsmSDWA<HasDst, NumSrcArgs, DstVT>.ret;
2165 field string AsmSDWA9 = getAsmSDWA9<HasDst, HasSDWAOMod, NumSrcArgs, DstVT>.ret;
2167 field string TieRegDPP = "$old";
2170 class VOP_NO_EXT <VOPProfile p> : VOPProfile <p.ArgVT> {
2174 let HasExtSDWA9 = 0;
2177 class VOP_PAT_GEN <VOPProfile p, int mode=PatGenMode.Pattern> : VOPProfile <p.ArgVT> {
2178 let NeedPatGen = mode;
2181 def VOP_F16_F16 : VOPProfile <[f16, f16, untyped, untyped]>;
2182 def VOP_F16_I16 : VOPProfile <[f16, i16, untyped, untyped]>;
2183 def VOP_I16_F16 : VOPProfile <[i16, f16, untyped, untyped]>;
2185 def VOP_F16_F16_F16 : VOPProfile <[f16, f16, f16, untyped]>;
2186 def VOP_F16_F16_I16 : VOPProfile <[f16, f16, i16, untyped]>;
2187 def VOP_F16_F16_I32 : VOPProfile <[f16, f16, i32, untyped]>;
2188 def VOP_I16_I16_I16 : VOPProfile <[i16, i16, i16, untyped]>;
2190 def VOP_I16_I16_I16_I16 : VOPProfile <[i16, i16, i16, i16, untyped]>;
2191 def VOP_F16_F16_F16_F16 : VOPProfile <[f16, f16, f16, f16, untyped]>;
2193 def VOP_I32_I16_I16_I32 : VOPProfile <[i32, i16, i16, i32, untyped]>;
2195 def VOP_V2F16_V2F16_V2F16 : VOPProfile <[v2f16, v2f16, v2f16, untyped]>;
2196 def VOP_V2I16_V2I16_V2I16 : VOPProfile <[v2i16, v2i16, v2i16, untyped]>;
2197 def VOP_B32_F16_F16 : VOPProfile <[i32, f16, f16, untyped]>;
2199 def VOP_V2F16_V2F16_V2F16_V2F16 : VOPProfile <[v2f16, v2f16, v2f16, v2f16]>;
2200 def VOP_V2I16_V2I16_V2I16_V2I16 : VOPProfile <[v2i16, v2i16, v2i16, v2i16]>;
2201 def VOP_V2I16_F32_F32 : VOPProfile <[v2i16, f32, f32, untyped]>;
2202 def VOP_V2I16_I32_I32 : VOPProfile <[v2i16, i32, i32, untyped]>;
2204 def VOP_F32_V2F16_V2F16_V2F16 : VOPProfile <[f32, v2f16, v2f16, v2f16]>;
2206 def VOP_NONE : VOPProfile <[untyped, untyped, untyped, untyped]>;
2208 def VOP_F32_F32 : VOPProfile <[f32, f32, untyped, untyped]>;
2209 def VOP_F32_F64 : VOPProfile <[f32, f64, untyped, untyped]>;
2210 def VOP_F32_I32 : VOPProfile <[f32, i32, untyped, untyped]>;
2211 def VOP_F64_F32 : VOPProfile <[f64, f32, untyped, untyped]>;
2212 def VOP_F64_F64 : VOPProfile <[f64, f64, untyped, untyped]>;
2213 def VOP_F64_I32 : VOPProfile <[f64, i32, untyped, untyped]>;
2214 def VOP_I32_F32 : VOPProfile <[i32, f32, untyped, untyped]>;
2215 def VOP_I32_F64 : VOPProfile <[i32, f64, untyped, untyped]>;
2216 def VOP_I32_I32 : VOPProfile <[i32, i32, untyped, untyped]>;
2217 def VOP_F16_F32 : VOPProfile <[f16, f32, untyped, untyped]>;
2218 def VOP_F32_F16 : VOPProfile <[f32, f16, untyped, untyped]>;
2220 def VOP_F32_F32_F16 : VOPProfile <[f32, f32, f16, untyped]>;
2221 def VOP_F32_F32_F32 : VOPProfile <[f32, f32, f32, untyped]>;
2222 def VOP_F32_F32_I32 : VOPProfile <[f32, f32, i32, untyped]>;
2223 def VOP_F64_F64_F64 : VOPProfile <[f64, f64, f64, untyped]>;
2224 def VOP_F64_F64_I32 : VOPProfile <[f64, f64, i32, untyped]>;
2225 def VOP_I32_F32_F32 : VOPProfile <[i32, f32, f32, untyped]>;
2226 def VOP_I32_F32_I32 : VOPProfile <[i32, f32, i32, untyped]>;
2227 def VOP_I32_I32_I32 : VOPProfile <[i32, i32, i32, untyped]>;
2228 def VOP_I32_I32_I32_ARITH : VOPProfile <[i32, i32, i32, untyped], 0, /*EnableClamp=*/1>;
2229 def VOP_V2F16_F32_F32 : VOPProfile <[v2f16, f32, f32, untyped]>;
2230 def VOP_F32_F16_F16_F16 : VOPProfile <[f32, f16, f16, f16]>;
2232 def VOP_I64_I64_I32 : VOPProfile <[i64, i64, i32, untyped]>;
2233 def VOP_I64_I32_I64 : VOPProfile <[i64, i32, i64, untyped]>;
2234 def VOP_I64_I64_I64 : VOPProfile <[i64, i64, i64, untyped]>;
2236 def VOP_F16_F32_F16_F32 : VOPProfile <[f16, f32, f16, f32]>;
2237 def VOP_F32_F32_F16_F16 : VOPProfile <[f32, f32, f16, f16]>;
2238 def VOP_F32_F32_F32_F32 : VOPProfile <[f32, f32, f32, f32]>;
2239 def VOP_F64_F64_F64_F64 : VOPProfile <[f64, f64, f64, f64]>;
2240 def VOP_I32_I32_I32_I32 : VOPProfile <[i32, i32, i32, i32]>;
2241 def VOP_I64_I32_I32_I64 : VOPProfile <[i64, i32, i32, i64]>;
2242 def VOP_I32_F32_I32_I32 : VOPProfile <[i32, f32, i32, i32]>;
2243 def VOP_I64_I64_I32_I64 : VOPProfile <[i64, i64, i32, i64]>;
2244 def VOP_V4I32_I64_I32_V4I32 : VOPProfile <[v4i32, i64, i32, v4i32]>;
2246 def VOP_F32_V2F16_V2F16_F32 : VOPProfile <[f32, v2f16, v2f16, f32]>;
2247 def VOP_I32_V2I16_V2I16_I32 : VOPProfile <[i32, v2i16, v2i16, i32]>;
2249 def VOP_V4F32_F32_F32_V4F32 : VOPProfile <[v4f32, f32, f32, v4f32]>;
2250 def VOP_V16F32_F32_F32_V16F32 : VOPProfile <[v16f32, f32, f32, v16f32]>;
2251 def VOP_V32F32_F32_F32_V32F32 : VOPProfile <[v32f32, f32, f32, v32f32]>;
2252 def VOP_V4F32_V4F16_V4F16_V4F32 : VOPProfile <[v4f32, v4f16, v4f16, v4f32]>;
2253 def VOP_V16F32_V4F16_V4F16_V16F32 : VOPProfile <[v16f32, v4f16, v4f16, v16f32]>;
2254 def VOP_V32F32_V4F16_V4F16_V32F32 : VOPProfile <[v32f32, v4f16, v4f16, v32f32]>;
2255 def VOP_V4F32_V2I16_V2I16_V4F32 : VOPProfile <[v4f32, v2i16, v2i16, v4f32]>;
2256 def VOP_V16F32_V2I16_V2I16_V16F32 : VOPProfile <[v16f32, v2i16, v2i16, v16f32]>;
2257 def VOP_V32F32_V2I16_V2I16_V32F32 : VOPProfile <[v32f32, v2i16, v2i16, v32f32]>;
2258 def VOP_V4I32_I32_I32_V4I32 : VOPProfile <[v4i32, i32, i32, v4i32]>;
2259 def VOP_V16I32_I32_I32_V16I32 : VOPProfile <[v16i32, i32, i32, v16i32]>;
2260 def VOP_V32I32_I32_I32_V32I32 : VOPProfile <[v32i32, i32, i32, v32i32]>;
2262 class Commutable_REV <string revOp, bit isOrig> {
2263 string RevOp = revOp;
2264 bit IsOrig = isOrig;
2267 class AtomicNoRet <string noRetOp, bit isRet> {
2268 string NoRetOp = noRetOp;
2272 //===----------------------------------------------------------------------===//
2273 // Interpolation opcodes
2274 //===----------------------------------------------------------------------===//
2276 class VINTRPDstOperand <RegisterClass rc> : RegisterOperand <rc, "printVINTRPDst">;
2278 class VINTRP_Pseudo <string opName, dag outs, dag ins, list<dag> pattern> :
2279 VINTRPCommon <outs, ins, "", pattern>,
2280 SIMCInstr<opName, SIEncodingFamily.NONE> {
2282 let isCodeGenOnly = 1;
2285 // FIXME-GFX10: WIP.
2286 class VINTRP_Real_si <bits <2> op, string opName, dag outs, dag ins,
2287 string asm, int encodingFamily> :
2288 VINTRPCommon <outs, ins, asm, []>,
2290 SIMCInstr<opName, encodingFamily> {
2291 let DisableDecoder = DisableSIDecoder;
2294 class VINTRP_Real_vi <bits <2> op, string opName, dag outs, dag ins,
2296 VINTRPCommon <outs, ins, asm, []>,
2298 SIMCInstr<opName, SIEncodingFamily.VI> {
2299 let AssemblerPredicate = VIAssemblerPredicate;
2300 let DecoderNamespace = "GFX8";
2301 let DisableDecoder = DisableVIDecoder;
2304 // FIXME-GFX10: WIP.
2305 multiclass VINTRP_m <bits <2> op, dag outs, dag ins, string asm,
2306 list<dag> pattern = []> {
2307 def "" : VINTRP_Pseudo <NAME, outs, ins, pattern>;
2309 let AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" in {
2310 def _si : VINTRP_Real_si <op, NAME, outs, ins, asm, SIEncodingFamily.SI>;
2311 } // End AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7"
2313 def _vi : VINTRP_Real_vi <op, NAME, outs, ins, asm>;
2315 let AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" in {
2316 def _gfx10 : VINTRP_Real_si<op, NAME, outs, ins, asm, SIEncodingFamily.GFX10>;
2317 } // End AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10"
2319 //===----------------------------------------------------------------------===//
2320 // Vector instruction mappings
2321 //===----------------------------------------------------------------------===//
2323 // Maps an opcode in e32 form to its e64 equivalent
2324 def getVOPe64 : InstrMapping {
2325 let FilterClass = "VOP";
2326 let RowFields = ["OpName"];
2327 let ColFields = ["Size", "VOP3"];
2328 let KeyCol = ["4", "0"];
2329 let ValueCols = [["8", "1"]];
2332 // Maps an opcode in e64 form to its e32 equivalent
2333 def getVOPe32 : InstrMapping {
2334 let FilterClass = "VOP";
2335 let RowFields = ["OpName"];
2336 let ColFields = ["Size", "VOP3"];
2337 let KeyCol = ["8", "1"];
2338 let ValueCols = [["4", "0"]];
2341 // Maps ordinary instructions to their SDWA counterparts
2342 def getSDWAOp : InstrMapping {
2343 let FilterClass = "VOP";
2344 let RowFields = ["OpName"];
2345 let ColFields = ["AsmVariantName"];
2346 let KeyCol = ["Default"];
2347 let ValueCols = [["SDWA"]];
2350 // Maps SDWA instructions to their ordinary counterparts
2351 def getBasicFromSDWAOp : InstrMapping {
2352 let FilterClass = "VOP";
2353 let RowFields = ["OpName"];
2354 let ColFields = ["AsmVariantName"];
2355 let KeyCol = ["SDWA"];
2356 let ValueCols = [["Default"]];
2359 // Maps ordinary instructions to their DPP counterparts
2360 def getDPPOp32 : InstrMapping {
2361 let FilterClass = "VOP";
2362 let RowFields = ["OpName"];
2363 let ColFields = ["AsmVariantName"];
2364 let KeyCol = ["Default"];
2365 let ValueCols = [["DPP"]];
2368 // Maps an commuted opcode to its original version
2369 def getCommuteOrig : InstrMapping {
2370 let FilterClass = "Commutable_REV";
2371 let RowFields = ["RevOp"];
2372 let ColFields = ["IsOrig"];
2374 let ValueCols = [["1"]];
2377 // Maps an original opcode to its commuted version
2378 def getCommuteRev : InstrMapping {
2379 let FilterClass = "Commutable_REV";
2380 let RowFields = ["RevOp"];
2381 let ColFields = ["IsOrig"];
2383 let ValueCols = [["0"]];
2386 def getMCOpcodeGen : InstrMapping {
2387 let FilterClass = "SIMCInstr";
2388 let RowFields = ["PseudoInstr"];
2389 let ColFields = ["Subtarget"];
2390 let KeyCol = [!cast<string>(SIEncodingFamily.NONE)];
2391 let ValueCols = [[!cast<string>(SIEncodingFamily.SI)],
2392 [!cast<string>(SIEncodingFamily.VI)],
2393 [!cast<string>(SIEncodingFamily.SDWA)],
2394 [!cast<string>(SIEncodingFamily.SDWA9)],
2395 // GFX80 encoding is added to work around a multiple matching
2396 // issue for buffer instructions with unpacked d16 data. This
2397 // does not actually change the encoding, and thus may be
2399 [!cast<string>(SIEncodingFamily.GFX80)],
2400 [!cast<string>(SIEncodingFamily.GFX9)],
2401 [!cast<string>(SIEncodingFamily.GFX10)],
2402 [!cast<string>(SIEncodingFamily.SDWA10)]];
2405 // Get equivalent SOPK instruction.
2406 def getSOPKOp : InstrMapping {
2407 let FilterClass = "SOPKInstTable";
2408 let RowFields = ["BaseCmpOp"];
2409 let ColFields = ["IsSOPK"];
2411 let ValueCols = [["1"]];
2414 def getAddr64Inst : InstrMapping {
2415 let FilterClass = "MUBUFAddr64Table";
2416 let RowFields = ["OpName"];
2417 let ColFields = ["IsAddr64"];
2419 let ValueCols = [["1"]];
2422 def getIfAddr64Inst : InstrMapping {
2423 let FilterClass = "MUBUFAddr64Table";
2424 let RowFields = ["OpName"];
2425 let ColFields = ["IsAddr64"];
2427 let ValueCols = [["1"]];
2430 def getMUBUFNoLdsInst : InstrMapping {
2431 let FilterClass = "MUBUFLdsTable";
2432 let RowFields = ["OpName"];
2433 let ColFields = ["IsLds"];
2435 let ValueCols = [["0"]];
2438 // Maps an atomic opcode to its version with a return value.
2439 def getAtomicRetOp : InstrMapping {
2440 let FilterClass = "AtomicNoRet";
2441 let RowFields = ["NoRetOp"];
2442 let ColFields = ["IsRet"];
2444 let ValueCols = [["1"]];
2447 // Maps an atomic opcode to its returnless version.
2448 def getAtomicNoRetOp : InstrMapping {
2449 let FilterClass = "AtomicNoRet";
2450 let RowFields = ["NoRetOp"];
2451 let ColFields = ["IsRet"];
2453 let ValueCols = [["0"]];
2456 // Maps a GLOBAL to its SADDR form.
2457 def getGlobalSaddrOp : InstrMapping {
2458 let FilterClass = "GlobalSaddrTable";
2459 let RowFields = ["SaddrOp"];
2460 let ColFields = ["IsSaddr"];
2462 let ValueCols = [["1"]];
2465 // Maps a v_cmpx opcode with sdst to opcode without sdst.
2466 def getVCMPXNoSDstOp : InstrMapping {
2467 let FilterClass = "VCMPXNoSDstTable";
2468 let RowFields = ["NoSDstOp"];
2469 let ColFields = ["HasSDst"];
2471 let ValueCols = [["0"]];
2474 // Maps a SOPP to a SOPP with S_NOP
2475 def getSOPPWithRelaxation : InstrMapping {
2476 let FilterClass = "Base_SOPP";
2477 let RowFields = ["AsmString"];
2478 let ColFields = ["Size"];
2480 let ValueCols = [["8"]];
2483 include "SIInstructions.td"
2485 include "DSInstructions.td"
2486 include "MIMGInstructions.td"