1 //===-- VOP1Instructions.td - Vector Instruction Defintions ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 //===----------------------------------------------------------------------===//
12 //===----------------------------------------------------------------------===//
14 class VOP1e <bits<8> op, VOPProfile P> : Enc32 {
18 let Inst{8-0} = !if(P.HasSrc0, src0{8-0}, 0);
20 let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
21 let Inst{31-25} = 0x3f; //encoding
24 class VOP1_SDWAe <bits<8> op, VOPProfile P> : VOP_SDWAe <P> {
27 let Inst{8-0} = 0xf9; // sdwa
29 let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
30 let Inst{31-25} = 0x3f; // encoding
33 class VOP1_SDWA9Ae <bits<8> op, VOPProfile P> : VOP_SDWA9Ae <P> {
36 let Inst{8-0} = 0xf9; // sdwa
38 let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
39 let Inst{31-25} = 0x3f; // encoding
42 class VOP1_Pseudo <string opName, VOPProfile P, list<dag> pattern=[], bit VOP1Only = 0> :
43 InstSI <P.Outs32, P.Ins32, "", pattern>,
45 SIMCInstr <!if(VOP1Only, opName, opName#"_e32"), SIEncodingFamily.NONE>,
46 MnemonicAlias<!if(VOP1Only, opName, opName#"_e32"), opName> {
49 let isCodeGenOnly = 1;
50 let UseNamedOperandTable = 1;
52 string Mnemonic = opName;
53 string AsmOperands = P.Asm32;
58 let hasSideEffects = 0;
59 let SubtargetPredicate = isGCN;
65 let AsmVariantName = AMDGPUAsmVariants.Default;
70 class VOP1_Real <VOP1_Pseudo ps, int EncodingFamily> :
71 InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []>,
72 SIMCInstr <ps.PseudoInstr, EncodingFamily> {
75 let isCodeGenOnly = 0;
77 let Constraints = ps.Constraints;
78 let DisableEncoding = ps.DisableEncoding;
80 // copy relevant pseudo op flags
81 let SubtargetPredicate = ps.SubtargetPredicate;
82 let AsmMatchConverter = ps.AsmMatchConverter;
83 let AsmVariantName = ps.AsmVariantName;
84 let Constraints = ps.Constraints;
85 let DisableEncoding = ps.DisableEncoding;
86 let TSFlags = ps.TSFlags;
87 let UseNamedOperandTable = ps.UseNamedOperandTable;
91 class VOP1_SDWA_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
92 VOP_SDWA_Pseudo <OpName, P, pattern> {
93 let AsmMatchConverter = "cvtSdwaVOP1";
96 class VOP1_SDWA9_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
97 VOP_SDWA9_Pseudo <OpName, P, pattern> {
98 let AsmMatchConverter = "cvtSdwaVOP1";
101 class getVOP1Pat64 <SDPatternOperator node, VOPProfile P> : LetDummies {
104 [(set P.DstVT:$vdst, (node (P.Src0VT (VOP3Mods0 P.Src0VT:$src0,
106 i1:$clamp, i32:$omod))))],
108 [(set P.DstVT:$vdst, (node (P.Src0VT (VOP3OMods P.Src0VT:$src0,
109 i1:$clamp, i32:$omod))))],
110 [(set P.DstVT:$vdst, (node P.Src0VT:$src0))]
115 multiclass VOP1Inst <string opName, VOPProfile P,
116 SDPatternOperator node = null_frag> {
117 def _e32 : VOP1_Pseudo <opName, P>;
118 def _e64 : VOP3_Pseudo <opName, P, getVOP1Pat64<node, P>.ret>;
119 def _sdwa : VOP1_SDWA_Pseudo <opName, P>;
120 def _sdwa9 : VOP1_SDWA9_Pseudo <opName, P>;
123 // Special profile for instructions which have clamp
124 // and output modifiers (but have no input modifiers)
125 class VOPProfileI2F<ValueType dstVt, ValueType srcVt> :
126 VOPProfile<[dstVt, srcVt, untyped, untyped]> {
128 let Ins64 = (ins Src0RC64:$src0, clampmod:$clamp, omod:$omod);
129 let Asm64 = "$vdst, $src0$clamp$omod";
131 let HasModifiers = 0;
136 def VOP1_F64_I32 : VOPProfileI2F <f64, i32>;
137 def VOP1_F32_I32 : VOPProfileI2F <f32, i32>;
138 def VOP1_F16_I16 : VOPProfileI2F <f16, i16>;
140 //===----------------------------------------------------------------------===//
142 //===----------------------------------------------------------------------===//
144 let VOPAsmPrefer32Bit = 1 in {
145 defm V_NOP : VOP1Inst <"v_nop", VOP_NONE>;
148 let isMoveImm = 1, isReMaterializable = 1, isAsCheapAsAMove = 1 in {
149 defm V_MOV_B32 : VOP1Inst <"v_mov_b32", VOP_I32_I32>;
150 } // End isMoveImm = 1
152 // FIXME: Specify SchedRW for READFIRSTLANE_B32
153 // TODO: Make profile for this, there is VOP3 encoding also
154 def V_READFIRSTLANE_B32 :
155 InstSI <(outs SReg_32:$vdst),
157 "v_readfirstlane_b32 $vdst, $src0",
158 [(set i32:$vdst, (int_amdgcn_readfirstlane i32:$src0))]>,
161 let isCodeGenOnly = 0;
162 let UseNamedOperandTable = 1;
167 let hasSideEffects = 0;
168 let SubtargetPredicate = isGCN;
173 let isConvergent = 1;
178 let Inst{8-0} = src0;
179 let Inst{16-9} = 0x2;
180 let Inst{24-17} = vdst;
181 let Inst{31-25} = 0x3f; //encoding
184 let SchedRW = [WriteQuarterRate32] in {
185 defm V_CVT_I32_F64 : VOP1Inst <"v_cvt_i32_f64", VOP_I32_F64, fp_to_sint>;
186 defm V_CVT_F64_I32 : VOP1Inst <"v_cvt_f64_i32", VOP1_F64_I32, sint_to_fp>;
187 defm V_CVT_F32_I32 : VOP1Inst <"v_cvt_f32_i32", VOP1_F32_I32, sint_to_fp>;
188 defm V_CVT_F32_U32 : VOP1Inst <"v_cvt_f32_u32", VOP1_F32_I32, uint_to_fp>;
189 defm V_CVT_U32_F32 : VOP1Inst <"v_cvt_u32_f32", VOP_I32_F32, fp_to_uint>;
190 defm V_CVT_I32_F32 : VOP1Inst <"v_cvt_i32_f32", VOP_I32_F32, fp_to_sint>;
191 defm V_CVT_F16_F32 : VOP1Inst <"v_cvt_f16_f32", VOP_F16_F32, fpround>;
192 defm V_CVT_F32_F16 : VOP1Inst <"v_cvt_f32_f16", VOP_F32_F16, fpextend>;
193 defm V_CVT_RPI_I32_F32 : VOP1Inst <"v_cvt_rpi_i32_f32", VOP_I32_F32, cvt_rpi_i32_f32>;
194 defm V_CVT_FLR_I32_F32 : VOP1Inst <"v_cvt_flr_i32_f32", VOP_I32_F32, cvt_flr_i32_f32>;
195 defm V_CVT_OFF_F32_I4 : VOP1Inst <"v_cvt_off_f32_i4", VOP1_F32_I32>;
196 defm V_CVT_F32_F64 : VOP1Inst <"v_cvt_f32_f64", VOP_F32_F64, fpround>;
197 defm V_CVT_F64_F32 : VOP1Inst <"v_cvt_f64_f32", VOP_F64_F32, fpextend>;
198 defm V_CVT_F32_UBYTE0 : VOP1Inst <"v_cvt_f32_ubyte0", VOP1_F32_I32, AMDGPUcvt_f32_ubyte0>;
199 defm V_CVT_F32_UBYTE1 : VOP1Inst <"v_cvt_f32_ubyte1", VOP1_F32_I32, AMDGPUcvt_f32_ubyte1>;
200 defm V_CVT_F32_UBYTE2 : VOP1Inst <"v_cvt_f32_ubyte2", VOP1_F32_I32, AMDGPUcvt_f32_ubyte2>;
201 defm V_CVT_F32_UBYTE3 : VOP1Inst <"v_cvt_f32_ubyte3", VOP1_F32_I32, AMDGPUcvt_f32_ubyte3>;
202 defm V_CVT_U32_F64 : VOP1Inst <"v_cvt_u32_f64", VOP_I32_F64, fp_to_uint>;
203 defm V_CVT_F64_U32 : VOP1Inst <"v_cvt_f64_u32", VOP1_F64_I32, uint_to_fp>;
204 } // End SchedRW = [WriteQuarterRate32]
206 defm V_FRACT_F32 : VOP1Inst <"v_fract_f32", VOP_F32_F32, AMDGPUfract>;
207 defm V_TRUNC_F32 : VOP1Inst <"v_trunc_f32", VOP_F32_F32, ftrunc>;
208 defm V_CEIL_F32 : VOP1Inst <"v_ceil_f32", VOP_F32_F32, fceil>;
209 defm V_RNDNE_F32 : VOP1Inst <"v_rndne_f32", VOP_F32_F32, frint>;
210 defm V_FLOOR_F32 : VOP1Inst <"v_floor_f32", VOP_F32_F32, ffloor>;
211 defm V_EXP_F32 : VOP1Inst <"v_exp_f32", VOP_F32_F32, fexp2>;
213 let SchedRW = [WriteQuarterRate32] in {
214 defm V_LOG_F32 : VOP1Inst <"v_log_f32", VOP_F32_F32, flog2>;
215 defm V_RCP_F32 : VOP1Inst <"v_rcp_f32", VOP_F32_F32, AMDGPUrcp>;
216 defm V_RCP_IFLAG_F32 : VOP1Inst <"v_rcp_iflag_f32", VOP_F32_F32>;
217 defm V_RSQ_F32 : VOP1Inst <"v_rsq_f32", VOP_F32_F32, AMDGPUrsq>;
218 } // End SchedRW = [WriteQuarterRate32]
220 let SchedRW = [WriteDouble] in {
221 defm V_RCP_F64 : VOP1Inst <"v_rcp_f64", VOP_F64_F64, AMDGPUrcp>;
222 defm V_RSQ_F64 : VOP1Inst <"v_rsq_f64", VOP_F64_F64, AMDGPUrsq>;
223 } // End SchedRW = [WriteDouble];
225 defm V_SQRT_F32 : VOP1Inst <"v_sqrt_f32", VOP_F32_F32, fsqrt>;
227 let SchedRW = [WriteDouble] in {
228 defm V_SQRT_F64 : VOP1Inst <"v_sqrt_f64", VOP_F64_F64, fsqrt>;
229 } // End SchedRW = [WriteDouble]
231 let SchedRW = [WriteQuarterRate32] in {
232 defm V_SIN_F32 : VOP1Inst <"v_sin_f32", VOP_F32_F32, AMDGPUsin>;
233 defm V_COS_F32 : VOP1Inst <"v_cos_f32", VOP_F32_F32, AMDGPUcos>;
234 } // End SchedRW = [WriteQuarterRate32]
236 defm V_NOT_B32 : VOP1Inst <"v_not_b32", VOP_I32_I32>;
237 defm V_BFREV_B32 : VOP1Inst <"v_bfrev_b32", VOP_I32_I32>;
238 defm V_FFBH_U32 : VOP1Inst <"v_ffbh_u32", VOP_I32_I32>;
239 defm V_FFBL_B32 : VOP1Inst <"v_ffbl_b32", VOP_I32_I32>;
240 defm V_FFBH_I32 : VOP1Inst <"v_ffbh_i32", VOP_I32_I32>;
241 defm V_FREXP_EXP_I32_F64 : VOP1Inst <"v_frexp_exp_i32_f64", VOP_I32_F64, int_amdgcn_frexp_exp>;
243 let SchedRW = [WriteDoubleAdd] in {
244 defm V_FREXP_MANT_F64 : VOP1Inst <"v_frexp_mant_f64", VOP_F64_F64, int_amdgcn_frexp_mant>;
245 defm V_FRACT_F64 : VOP1Inst <"v_fract_f64", VOP_F64_F64, AMDGPUfract>;
246 } // End SchedRW = [WriteDoubleAdd]
248 defm V_FREXP_EXP_I32_F32 : VOP1Inst <"v_frexp_exp_i32_f32", VOP_I32_F32, int_amdgcn_frexp_exp>;
249 defm V_FREXP_MANT_F32 : VOP1Inst <"v_frexp_mant_f32", VOP_F32_F32, int_amdgcn_frexp_mant>;
251 let VOPAsmPrefer32Bit = 1 in {
252 defm V_CLREXCP : VOP1Inst <"v_clrexcp", VOP_NO_EXT<VOP_NONE>>;
255 // Restrict src0 to be VGPR
256 def VOP_I32_VI32_NO_EXT : VOPProfile<[i32, i32, untyped, untyped]> {
257 let Src0RC32 = VRegSrc_32;
258 let Src0RC64 = VRegSrc_32;
264 // Special case because there are no true output operands. Hack vdst
265 // to be a src operand. The custom inserter must add a tied implicit
266 // def and use of the super register since there seems to be no way to
267 // add an implicit def of a virtual register in tablegen.
268 def VOP_MOVRELD : VOPProfile<[untyped, i32, untyped, untyped]> {
269 let Src0RC32 = VOPDstOperand<VGPR_32>;
270 let Src0RC64 = VOPDstOperand<VGPR_32>;
273 let Ins32 = (ins Src0RC32:$vdst, VSrc_b32:$src0);
274 let Ins64 = (ins Src0RC64:$vdst, VSrc_b32:$src0);
275 let InsDPP = (ins Src0RC32:$vdst, Src0RC32:$src0, dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
276 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl);
277 let InsSDWA = (ins Src0RC32:$vdst, Src0ModSDWA:$src0_modifiers, Src0SDWA:$src0,
278 clampmod:$clamp, dst_sel:$dst_sel, dst_unused:$dst_unused,
280 let InsSDWA9 = (ins Src0RC32:$vdst, Src0ModSDWA9:$src0_modifiers, Src0SDWA9:$src0,
281 clampmod:$clamp, omod:$omod, dst_sel:$dst_sel, dst_unused:$dst_unused,
284 let Asm32 = getAsm32<1, 1>.ret;
285 let Asm64 = getAsm64<1, 1, 0, 1>.ret;
286 let AsmDPP = getAsmDPP<1, 1, 0>.ret;
287 let AsmSDWA = getAsmSDWA<1, 1>.ret;
288 let AsmSDWA9 = getAsmSDWA9<1, 0, 1>.ret;
293 let EmitDst = 1; // force vdst emission
296 let SubtargetPredicate = HasMovrel, Uses = [M0, EXEC] in {
297 // v_movreld_b32 is a special case because the destination output
298 // register is really a source. It isn't actually read (but may be
299 // written), and is only to provide the base register to start
300 // indexing from. Tablegen seems to not let you define an implicit
301 // virtual register output for the super register being written into,
302 // so this must have an implicit def of the register added to it.
303 defm V_MOVRELD_B32 : VOP1Inst <"v_movreld_b32", VOP_MOVRELD>;
304 defm V_MOVRELS_B32 : VOP1Inst <"v_movrels_b32", VOP_I32_VI32_NO_EXT>;
305 defm V_MOVRELSD_B32 : VOP1Inst <"v_movrelsd_b32", VOP_NO_EXT<VOP_I32_I32>>;
306 } // End Uses = [M0, EXEC]
308 let SchedRW = [WriteQuarterRate32] in {
309 defm V_MOV_FED_B32 : VOP1Inst <"v_mov_fed_b32", VOP_I32_I32>;
312 // These instruction only exist on SI and CI
313 let SubtargetPredicate = isSICI in {
315 let SchedRW = [WriteQuarterRate32] in {
316 defm V_LOG_CLAMP_F32 : VOP1Inst <"v_log_clamp_f32", VOP_F32_F32, int_amdgcn_log_clamp>;
317 defm V_RCP_CLAMP_F32 : VOP1Inst <"v_rcp_clamp_f32", VOP_F32_F32>;
318 defm V_RCP_LEGACY_F32 : VOP1Inst <"v_rcp_legacy_f32", VOP_F32_F32, AMDGPUrcp_legacy>;
319 defm V_RSQ_CLAMP_F32 : VOP1Inst <"v_rsq_clamp_f32", VOP_F32_F32, AMDGPUrsq_clamp>;
320 defm V_RSQ_LEGACY_F32 : VOP1Inst <"v_rsq_legacy_f32", VOP_F32_F32, AMDGPUrsq_legacy>;
321 } // End SchedRW = [WriteQuarterRate32]
323 let SchedRW = [WriteDouble] in {
324 defm V_RCP_CLAMP_F64 : VOP1Inst <"v_rcp_clamp_f64", VOP_F64_F64>;
325 defm V_RSQ_CLAMP_F64 : VOP1Inst <"v_rsq_clamp_f64", VOP_F64_F64, AMDGPUrsq_clamp>;
326 } // End SchedRW = [WriteDouble]
328 } // End SubtargetPredicate = isSICI
331 let SubtargetPredicate = isCIVI in {
333 let SchedRW = [WriteDoubleAdd] in {
334 defm V_TRUNC_F64 : VOP1Inst <"v_trunc_f64", VOP_F64_F64, ftrunc>;
335 defm V_CEIL_F64 : VOP1Inst <"v_ceil_f64", VOP_F64_F64, fceil>;
336 defm V_FLOOR_F64 : VOP1Inst <"v_floor_f64", VOP_F64_F64, ffloor>;
337 defm V_RNDNE_F64 : VOP1Inst <"v_rndne_f64", VOP_F64_F64, frint>;
338 } // End SchedRW = [WriteDoubleAdd]
340 let SchedRW = [WriteQuarterRate32] in {
341 defm V_LOG_LEGACY_F32 : VOP1Inst <"v_log_legacy_f32", VOP_F32_F32>;
342 defm V_EXP_LEGACY_F32 : VOP1Inst <"v_exp_legacy_f32", VOP_F32_F32>;
343 } // End SchedRW = [WriteQuarterRate32]
345 } // End SubtargetPredicate = isCIVI
348 let SubtargetPredicate = Has16BitInsts in {
350 defm V_CVT_F16_U16 : VOP1Inst <"v_cvt_f16_u16", VOP1_F16_I16, uint_to_fp>;
351 defm V_CVT_F16_I16 : VOP1Inst <"v_cvt_f16_i16", VOP1_F16_I16, sint_to_fp>;
352 defm V_CVT_U16_F16 : VOP1Inst <"v_cvt_u16_f16", VOP_I16_F16, fp_to_uint>;
353 defm V_CVT_I16_F16 : VOP1Inst <"v_cvt_i16_f16", VOP_I16_F16, fp_to_sint>;
354 defm V_RCP_F16 : VOP1Inst <"v_rcp_f16", VOP_F16_F16, AMDGPUrcp>;
355 defm V_SQRT_F16 : VOP1Inst <"v_sqrt_f16", VOP_F16_F16, fsqrt>;
356 defm V_RSQ_F16 : VOP1Inst <"v_rsq_f16", VOP_F16_F16, AMDGPUrsq>;
357 defm V_LOG_F16 : VOP1Inst <"v_log_f16", VOP_F16_F16, flog2>;
358 defm V_EXP_F16 : VOP1Inst <"v_exp_f16", VOP_F16_F16, fexp2>;
359 defm V_FREXP_MANT_F16 : VOP1Inst <"v_frexp_mant_f16", VOP_F16_F16, int_amdgcn_frexp_mant>;
360 defm V_FREXP_EXP_I16_F16 : VOP1Inst <"v_frexp_exp_i16_f16", VOP_I16_F16, int_amdgcn_frexp_exp>;
361 defm V_FLOOR_F16 : VOP1Inst <"v_floor_f16", VOP_F16_F16, ffloor>;
362 defm V_CEIL_F16 : VOP1Inst <"v_ceil_f16", VOP_F16_F16, fceil>;
363 defm V_TRUNC_F16 : VOP1Inst <"v_trunc_f16", VOP_F16_F16, ftrunc>;
364 defm V_RNDNE_F16 : VOP1Inst <"v_rndne_f16", VOP_F16_F16, frint>;
365 defm V_FRACT_F16 : VOP1Inst <"v_fract_f16", VOP_F16_F16, AMDGPUfract>;
366 defm V_SIN_F16 : VOP1Inst <"v_sin_f16", VOP_F16_F16, AMDGPUsin>;
367 defm V_COS_F16 : VOP1Inst <"v_cos_f16", VOP_F16_F16, AMDGPUcos>;
371 let Predicates = [Has16BitInsts] in {
374 (f32 (f16_to_fp i16:$src)),
375 (V_CVT_F32_F16_e32 $src)
379 (i16 (AMDGPUfp_to_f16 f32:$src)),
380 (V_CVT_F16_F32_e32 $src)
385 def VOP_SWAP_I32 : VOPProfile<[i32, i32, i32, untyped]> {
386 let Outs32 = (outs VGPR_32:$vdst, VGPR_32:$vdst1);
387 let Ins32 = (ins VGPR_32:$src0, VGPR_32:$src1);
389 let Asm32 = " $vdst, $src0";
394 let SubtargetPredicate = isGFX9 in {
395 let Constraints = "$vdst = $src1, $vdst1 = $src0",
396 DisableEncoding="$vdst1,$src1",
397 SchedRW = [Write64Bit, Write64Bit] in {
398 // Never VOP3. Takes as long as 2 v_mov_b32s
399 def V_SWAP_B32 : VOP1_Pseudo <"v_swap_b32", VOP_SWAP_I32, [], 1>;
402 } // End SubtargetPredicate = isGFX9
404 //===----------------------------------------------------------------------===//
406 //===----------------------------------------------------------------------===//
408 //===----------------------------------------------------------------------===//
410 //===----------------------------------------------------------------------===//
412 multiclass VOP1_Real_si <bits<9> op> {
413 let AssemblerPredicates = [isSICI], DecoderNamespace = "SICI" in {
415 VOP1_Real<!cast<VOP1_Pseudo>(NAME#"_e32"), SIEncodingFamily.SI>,
416 VOP1e<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32").Pfl>;
418 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>,
419 VOP3e_si <{1, 1, op{6-0}}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
423 defm V_NOP : VOP1_Real_si <0x0>;
424 defm V_MOV_B32 : VOP1_Real_si <0x1>;
425 defm V_CVT_I32_F64 : VOP1_Real_si <0x3>;
426 defm V_CVT_F64_I32 : VOP1_Real_si <0x4>;
427 defm V_CVT_F32_I32 : VOP1_Real_si <0x5>;
428 defm V_CVT_F32_U32 : VOP1_Real_si <0x6>;
429 defm V_CVT_U32_F32 : VOP1_Real_si <0x7>;
430 defm V_CVT_I32_F32 : VOP1_Real_si <0x8>;
431 defm V_MOV_FED_B32 : VOP1_Real_si <0x9>;
432 defm V_CVT_F16_F32 : VOP1_Real_si <0xa>;
433 defm V_CVT_F32_F16 : VOP1_Real_si <0xb>;
434 defm V_CVT_RPI_I32_F32 : VOP1_Real_si <0xc>;
435 defm V_CVT_FLR_I32_F32 : VOP1_Real_si <0xd>;
436 defm V_CVT_OFF_F32_I4 : VOP1_Real_si <0xe>;
437 defm V_CVT_F32_F64 : VOP1_Real_si <0xf>;
438 defm V_CVT_F64_F32 : VOP1_Real_si <0x10>;
439 defm V_CVT_F32_UBYTE0 : VOP1_Real_si <0x11>;
440 defm V_CVT_F32_UBYTE1 : VOP1_Real_si <0x12>;
441 defm V_CVT_F32_UBYTE2 : VOP1_Real_si <0x13>;
442 defm V_CVT_F32_UBYTE3 : VOP1_Real_si <0x14>;
443 defm V_CVT_U32_F64 : VOP1_Real_si <0x15>;
444 defm V_CVT_F64_U32 : VOP1_Real_si <0x16>;
445 defm V_FRACT_F32 : VOP1_Real_si <0x20>;
446 defm V_TRUNC_F32 : VOP1_Real_si <0x21>;
447 defm V_CEIL_F32 : VOP1_Real_si <0x22>;
448 defm V_RNDNE_F32 : VOP1_Real_si <0x23>;
449 defm V_FLOOR_F32 : VOP1_Real_si <0x24>;
450 defm V_EXP_F32 : VOP1_Real_si <0x25>;
451 defm V_LOG_CLAMP_F32 : VOP1_Real_si <0x26>;
452 defm V_LOG_F32 : VOP1_Real_si <0x27>;
453 defm V_RCP_CLAMP_F32 : VOP1_Real_si <0x28>;
454 defm V_RCP_LEGACY_F32 : VOP1_Real_si <0x29>;
455 defm V_RCP_F32 : VOP1_Real_si <0x2a>;
456 defm V_RCP_IFLAG_F32 : VOP1_Real_si <0x2b>;
457 defm V_RSQ_CLAMP_F32 : VOP1_Real_si <0x2c>;
458 defm V_RSQ_LEGACY_F32 : VOP1_Real_si <0x2d>;
459 defm V_RSQ_F32 : VOP1_Real_si <0x2e>;
460 defm V_RCP_F64 : VOP1_Real_si <0x2f>;
461 defm V_RCP_CLAMP_F64 : VOP1_Real_si <0x30>;
462 defm V_RSQ_F64 : VOP1_Real_si <0x31>;
463 defm V_RSQ_CLAMP_F64 : VOP1_Real_si <0x32>;
464 defm V_SQRT_F32 : VOP1_Real_si <0x33>;
465 defm V_SQRT_F64 : VOP1_Real_si <0x34>;
466 defm V_SIN_F32 : VOP1_Real_si <0x35>;
467 defm V_COS_F32 : VOP1_Real_si <0x36>;
468 defm V_NOT_B32 : VOP1_Real_si <0x37>;
469 defm V_BFREV_B32 : VOP1_Real_si <0x38>;
470 defm V_FFBH_U32 : VOP1_Real_si <0x39>;
471 defm V_FFBL_B32 : VOP1_Real_si <0x3a>;
472 defm V_FFBH_I32 : VOP1_Real_si <0x3b>;
473 defm V_FREXP_EXP_I32_F64 : VOP1_Real_si <0x3c>;
474 defm V_FREXP_MANT_F64 : VOP1_Real_si <0x3d>;
475 defm V_FRACT_F64 : VOP1_Real_si <0x3e>;
476 defm V_FREXP_EXP_I32_F32 : VOP1_Real_si <0x3f>;
477 defm V_FREXP_MANT_F32 : VOP1_Real_si <0x40>;
478 defm V_CLREXCP : VOP1_Real_si <0x41>;
479 defm V_MOVRELD_B32 : VOP1_Real_si <0x42>;
480 defm V_MOVRELS_B32 : VOP1_Real_si <0x43>;
481 defm V_MOVRELSD_B32 : VOP1_Real_si <0x44>;
483 //===----------------------------------------------------------------------===//
485 //===----------------------------------------------------------------------===//
487 multiclass VOP1_Real_ci <bits<9> op> {
488 let AssemblerPredicates = [isCIOnly], DecoderNamespace = "CI" in {
490 VOP1_Real<!cast<VOP1_Pseudo>(NAME#"_e32"), SIEncodingFamily.SI>,
491 VOP1e<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32").Pfl>;
493 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>,
494 VOP3e_si <{1, 1, op{6-0}}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
498 defm V_TRUNC_F64 : VOP1_Real_ci <0x17>;
499 defm V_CEIL_F64 : VOP1_Real_ci <0x18>;
500 defm V_FLOOR_F64 : VOP1_Real_ci <0x1A>;
501 defm V_RNDNE_F64 : VOP1_Real_ci <0x19>;
502 defm V_LOG_LEGACY_F32 : VOP1_Real_ci <0x45>;
503 defm V_EXP_LEGACY_F32 : VOP1_Real_ci <0x46>;
505 //===----------------------------------------------------------------------===//
507 //===----------------------------------------------------------------------===//
509 class VOP1_DPP <bits<8> op, VOP1_Pseudo ps, VOPProfile P = ps.Pfl> :
510 VOP_DPP <ps.OpName, P> {
513 let SchedRW = ps.SchedRW;
514 let hasSideEffects = ps.hasSideEffects;
515 let Constraints = ps.Constraints;
516 let DisableEncoding = ps.DisableEncoding;
519 let Inst{8-0} = 0xfa; // dpp
521 let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
522 let Inst{31-25} = 0x3f; //encoding
525 multiclass VOP1Only_Real_vi <bits<10> op> {
526 let AssemblerPredicates = [isVI], DecoderNamespace = "VI" in {
528 VOP1_Real<!cast<VOP1_Pseudo>(NAME), SIEncodingFamily.VI>,
529 VOP1e<op{7-0}, !cast<VOP1_Pseudo>(NAME).Pfl>;
533 multiclass VOP1_Real_vi <bits<10> op> {
534 let AssemblerPredicates = [isVI], DecoderNamespace = "VI" in {
536 VOP1_Real<!cast<VOP1_Pseudo>(NAME#"_e32"), SIEncodingFamily.VI>,
537 VOP1e<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32").Pfl>;
539 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
540 VOP3e_vi <!add(0x140, op), !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
544 VOP_SDWA_Real <!cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa")>,
545 VOP1_SDWAe <op{7-0}, !cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa").Pfl>;
548 VOP_SDWA9_Real <!cast<VOP1_SDWA9_Pseudo>(NAME#"_sdwa9")>,
549 VOP1_SDWA9Ae <op{7-0}, !cast<VOP1_SDWA9_Pseudo>(NAME#"_sdwa9").Pfl>;
551 // For now left dpp only for asm/dasm
552 // TODO: add corresponding pseudo
553 def _dpp : VOP1_DPP<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32")>;
556 defm V_NOP : VOP1_Real_vi <0x0>;
557 defm V_MOV_B32 : VOP1_Real_vi <0x1>;
558 defm V_CVT_I32_F64 : VOP1_Real_vi <0x3>;
559 defm V_CVT_F64_I32 : VOP1_Real_vi <0x4>;
560 defm V_CVT_F32_I32 : VOP1_Real_vi <0x5>;
561 defm V_CVT_F32_U32 : VOP1_Real_vi <0x6>;
562 defm V_CVT_U32_F32 : VOP1_Real_vi <0x7>;
563 defm V_CVT_I32_F32 : VOP1_Real_vi <0x8>;
564 defm V_MOV_FED_B32 : VOP1_Real_vi <0x9>;
565 defm V_CVT_F16_F32 : VOP1_Real_vi <0xa>;
566 defm V_CVT_F32_F16 : VOP1_Real_vi <0xb>;
567 defm V_CVT_RPI_I32_F32 : VOP1_Real_vi <0xc>;
568 defm V_CVT_FLR_I32_F32 : VOP1_Real_vi <0xd>;
569 defm V_CVT_OFF_F32_I4 : VOP1_Real_vi <0xe>;
570 defm V_CVT_F32_F64 : VOP1_Real_vi <0xf>;
571 defm V_CVT_F64_F32 : VOP1_Real_vi <0x10>;
572 defm V_CVT_F32_UBYTE0 : VOP1_Real_vi <0x11>;
573 defm V_CVT_F32_UBYTE1 : VOP1_Real_vi <0x12>;
574 defm V_CVT_F32_UBYTE2 : VOP1_Real_vi <0x13>;
575 defm V_CVT_F32_UBYTE3 : VOP1_Real_vi <0x14>;
576 defm V_CVT_U32_F64 : VOP1_Real_vi <0x15>;
577 defm V_CVT_F64_U32 : VOP1_Real_vi <0x16>;
578 defm V_FRACT_F32 : VOP1_Real_vi <0x1b>;
579 defm V_TRUNC_F32 : VOP1_Real_vi <0x1c>;
580 defm V_CEIL_F32 : VOP1_Real_vi <0x1d>;
581 defm V_RNDNE_F32 : VOP1_Real_vi <0x1e>;
582 defm V_FLOOR_F32 : VOP1_Real_vi <0x1f>;
583 defm V_EXP_F32 : VOP1_Real_vi <0x20>;
584 defm V_LOG_F32 : VOP1_Real_vi <0x21>;
585 defm V_RCP_F32 : VOP1_Real_vi <0x22>;
586 defm V_RCP_IFLAG_F32 : VOP1_Real_vi <0x23>;
587 defm V_RSQ_F32 : VOP1_Real_vi <0x24>;
588 defm V_RCP_F64 : VOP1_Real_vi <0x25>;
589 defm V_RSQ_F64 : VOP1_Real_vi <0x26>;
590 defm V_SQRT_F32 : VOP1_Real_vi <0x27>;
591 defm V_SQRT_F64 : VOP1_Real_vi <0x28>;
592 defm V_SIN_F32 : VOP1_Real_vi <0x29>;
593 defm V_COS_F32 : VOP1_Real_vi <0x2a>;
594 defm V_NOT_B32 : VOP1_Real_vi <0x2b>;
595 defm V_BFREV_B32 : VOP1_Real_vi <0x2c>;
596 defm V_FFBH_U32 : VOP1_Real_vi <0x2d>;
597 defm V_FFBL_B32 : VOP1_Real_vi <0x2e>;
598 defm V_FFBH_I32 : VOP1_Real_vi <0x2f>;
599 defm V_FREXP_EXP_I32_F64 : VOP1_Real_vi <0x30>;
600 defm V_FREXP_MANT_F64 : VOP1_Real_vi <0x31>;
601 defm V_FRACT_F64 : VOP1_Real_vi <0x32>;
602 defm V_FREXP_EXP_I32_F32 : VOP1_Real_vi <0x33>;
603 defm V_FREXP_MANT_F32 : VOP1_Real_vi <0x34>;
604 defm V_CLREXCP : VOP1_Real_vi <0x35>;
605 defm V_MOVRELD_B32 : VOP1_Real_vi <0x36>;
606 defm V_MOVRELS_B32 : VOP1_Real_vi <0x37>;
607 defm V_MOVRELSD_B32 : VOP1_Real_vi <0x38>;
608 defm V_TRUNC_F64 : VOP1_Real_vi <0x17>;
609 defm V_CEIL_F64 : VOP1_Real_vi <0x18>;
610 defm V_FLOOR_F64 : VOP1_Real_vi <0x1A>;
611 defm V_RNDNE_F64 : VOP1_Real_vi <0x19>;
612 defm V_LOG_LEGACY_F32 : VOP1_Real_vi <0x4c>;
613 defm V_EXP_LEGACY_F32 : VOP1_Real_vi <0x4b>;
614 defm V_CVT_F16_U16 : VOP1_Real_vi <0x39>;
615 defm V_CVT_F16_I16 : VOP1_Real_vi <0x3a>;
616 defm V_CVT_U16_F16 : VOP1_Real_vi <0x3b>;
617 defm V_CVT_I16_F16 : VOP1_Real_vi <0x3c>;
618 defm V_RCP_F16 : VOP1_Real_vi <0x3d>;
619 defm V_SQRT_F16 : VOP1_Real_vi <0x3e>;
620 defm V_RSQ_F16 : VOP1_Real_vi <0x3f>;
621 defm V_LOG_F16 : VOP1_Real_vi <0x40>;
622 defm V_EXP_F16 : VOP1_Real_vi <0x41>;
623 defm V_FREXP_MANT_F16 : VOP1_Real_vi <0x42>;
624 defm V_FREXP_EXP_I16_F16 : VOP1_Real_vi <0x43>;
625 defm V_FLOOR_F16 : VOP1_Real_vi <0x44>;
626 defm V_CEIL_F16 : VOP1_Real_vi <0x45>;
627 defm V_TRUNC_F16 : VOP1_Real_vi <0x46>;
628 defm V_RNDNE_F16 : VOP1_Real_vi <0x47>;
629 defm V_FRACT_F16 : VOP1_Real_vi <0x48>;
630 defm V_SIN_F16 : VOP1_Real_vi <0x49>;
631 defm V_COS_F16 : VOP1_Real_vi <0x4a>;
632 defm V_SWAP_B32 : VOP1Only_Real_vi <0x51>;
634 // Copy of v_mov_b32 with $vdst as a use operand for use with VGPR
635 // indexing mode. vdst can't be treated as a def for codegen purposes,
636 // and an implicit use and def of the super register should be added.
637 def V_MOV_B32_indirect : VPseudoInstSI<(outs),
638 (ins getVALUDstForVT<i32>.ret:$vdst, getVOPSrc0ForVT<i32>.ret:$src0)>,
639 PseudoInstExpansion<(V_MOV_B32_e32_vi getVALUDstForVT<i32>.ret:$vdst,
640 getVOPSrc0ForVT<i32>.ret:$src0)> {
642 let SubtargetPredicate = isVI;
645 // This is a pseudo variant of the v_movreld_b32 instruction in which the
646 // vector operand appears only twice, once as def and once as use. Using this
647 // pseudo avoids problems with the Two Address instructions pass.
648 class V_MOVRELD_B32_pseudo<RegisterClass rc> : VPseudoInstSI <
650 (ins rc:$vsrc, VSrc_b32:$val, i32imm:$offset)> {
653 let Constraints = "$vsrc = $vdst";
654 let Uses = [M0, EXEC];
656 let SubtargetPredicate = HasMovrel;
659 def V_MOVRELD_B32_V1 : V_MOVRELD_B32_pseudo<VGPR_32>;
660 def V_MOVRELD_B32_V2 : V_MOVRELD_B32_pseudo<VReg_64>;
661 def V_MOVRELD_B32_V4 : V_MOVRELD_B32_pseudo<VReg_128>;
662 def V_MOVRELD_B32_V8 : V_MOVRELD_B32_pseudo<VReg_256>;
663 def V_MOVRELD_B32_V16 : V_MOVRELD_B32_pseudo<VReg_512>;
665 let Predicates = [isVI] in {
668 (i32 (int_amdgcn_mov_dpp i32:$src, imm:$dpp_ctrl, imm:$row_mask, imm:$bank_mask,
670 (V_MOV_B32_dpp $src, (as_i32imm $dpp_ctrl), (as_i32imm $row_mask),
671 (as_i32imm $bank_mask), (as_i1imm $bound_ctrl))
676 (i32 (anyext i16:$src)),
681 (i64 (anyext i16:$src)),
682 (REG_SEQUENCE VReg_64,
683 (i32 (COPY $src)), sub0,
684 (V_MOV_B32_e32 (i32 0)), sub1)
688 (i16 (trunc i32:$src)),
693 (i16 (trunc i64:$src)),
694 (EXTRACT_SUBREG $src, sub0)
697 } // End Predicates = [isVI]