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 VOP_Pseudo <opName, !if(VOP1Only, "", "_e32"), P, P.Outs32, P.Ins32, "", pattern> {
45 let AsmOperands = P.Asm32;
50 let hasSideEffects = 0;
51 let SubtargetPredicate = isGCN;
57 let AsmVariantName = AMDGPUAsmVariants.Default;
60 class VOP1_Real <VOP1_Pseudo ps, int EncodingFamily> :
61 InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []>,
62 SIMCInstr <ps.PseudoInstr, EncodingFamily> {
65 let isCodeGenOnly = 0;
67 let Constraints = ps.Constraints;
68 let DisableEncoding = ps.DisableEncoding;
70 // copy relevant pseudo op flags
71 let SubtargetPredicate = ps.SubtargetPredicate;
72 let AsmMatchConverter = ps.AsmMatchConverter;
73 let AsmVariantName = ps.AsmVariantName;
74 let Constraints = ps.Constraints;
75 let DisableEncoding = ps.DisableEncoding;
76 let TSFlags = ps.TSFlags;
77 let UseNamedOperandTable = ps.UseNamedOperandTable;
82 class VOP1_SDWA_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
83 VOP_SDWA_Pseudo <OpName, P, pattern> {
84 let AsmMatchConverter = "cvtSdwaVOP1";
87 class getVOP1Pat64 <SDPatternOperator node, VOPProfile P> : LetDummies {
90 [(set P.DstVT:$vdst, (node (P.Src0VT (VOP3Mods0 P.Src0VT:$src0,
92 i1:$clamp, i32:$omod))))],
94 [(set P.DstVT:$vdst, (node (P.Src0VT (VOP3OMods P.Src0VT:$src0,
95 i1:$clamp, i32:$omod))))],
96 [(set P.DstVT:$vdst, (node P.Src0VT:$src0))]
101 multiclass VOP1Inst <string opName, VOPProfile P,
102 SDPatternOperator node = null_frag> {
103 def _e32 : VOP1_Pseudo <opName, P>;
104 def _e64 : VOP3_Pseudo <opName, P, getVOP1Pat64<node, P>.ret>;
105 def _sdwa : VOP1_SDWA_Pseudo <opName, P>;
108 // Special profile for instructions which have clamp
109 // and output modifiers (but have no input modifiers)
110 class VOPProfileI2F<ValueType dstVt, ValueType srcVt> :
111 VOPProfile<[dstVt, srcVt, untyped, untyped]> {
113 let Ins64 = (ins Src0RC64:$src0, clampmod:$clamp, omod:$omod);
114 let Asm64 = "$vdst, $src0$clamp$omod";
116 let HasModifiers = 0;
121 def VOP1_F64_I32 : VOPProfileI2F <f64, i32>;
122 def VOP1_F32_I32 : VOPProfileI2F <f32, i32>;
123 def VOP1_F16_I16 : VOPProfileI2F <f16, i16>;
125 //===----------------------------------------------------------------------===//
127 //===----------------------------------------------------------------------===//
129 let VOPAsmPrefer32Bit = 1 in {
130 defm V_NOP : VOP1Inst <"v_nop", VOP_NONE>;
133 let isMoveImm = 1, isReMaterializable = 1, isAsCheapAsAMove = 1 in {
134 defm V_MOV_B32 : VOP1Inst <"v_mov_b32", VOP_I32_I32>;
135 } // End isMoveImm = 1
137 // FIXME: Specify SchedRW for READFIRSTLANE_B32
138 // TODO: Make profile for this, there is VOP3 encoding also
139 def V_READFIRSTLANE_B32 :
140 InstSI <(outs SReg_32:$vdst),
142 "v_readfirstlane_b32 $vdst, $src0",
143 [(set i32:$vdst, (int_amdgcn_readfirstlane i32:$src0))]>,
146 let isCodeGenOnly = 0;
147 let UseNamedOperandTable = 1;
152 let hasSideEffects = 0;
153 let SubtargetPredicate = isGCN;
158 let isConvergent = 1;
163 let Inst{8-0} = src0;
164 let Inst{16-9} = 0x2;
165 let Inst{24-17} = vdst;
166 let Inst{31-25} = 0x3f; //encoding
169 let SchedRW = [WriteQuarterRate32] in {
170 defm V_CVT_I32_F64 : VOP1Inst <"v_cvt_i32_f64", VOP_I32_F64, fp_to_sint>;
171 defm V_CVT_F64_I32 : VOP1Inst <"v_cvt_f64_i32", VOP1_F64_I32, sint_to_fp>;
172 defm V_CVT_F32_I32 : VOP1Inst <"v_cvt_f32_i32", VOP1_F32_I32, sint_to_fp>;
173 defm V_CVT_F32_U32 : VOP1Inst <"v_cvt_f32_u32", VOP1_F32_I32, uint_to_fp>;
174 defm V_CVT_U32_F32 : VOP1Inst <"v_cvt_u32_f32", VOP_I32_F32, fp_to_uint>;
175 defm V_CVT_I32_F32 : VOP1Inst <"v_cvt_i32_f32", VOP_I32_F32, fp_to_sint>;
176 defm V_CVT_F16_F32 : VOP1Inst <"v_cvt_f16_f32", VOP_F16_F32, fpround>;
177 defm V_CVT_F32_F16 : VOP1Inst <"v_cvt_f32_f16", VOP_F32_F16, fpextend>;
178 defm V_CVT_RPI_I32_F32 : VOP1Inst <"v_cvt_rpi_i32_f32", VOP_I32_F32, cvt_rpi_i32_f32>;
179 defm V_CVT_FLR_I32_F32 : VOP1Inst <"v_cvt_flr_i32_f32", VOP_I32_F32, cvt_flr_i32_f32>;
180 defm V_CVT_OFF_F32_I4 : VOP1Inst <"v_cvt_off_f32_i4", VOP1_F32_I32>;
181 defm V_CVT_F32_F64 : VOP1Inst <"v_cvt_f32_f64", VOP_F32_F64, fpround>;
182 defm V_CVT_F64_F32 : VOP1Inst <"v_cvt_f64_f32", VOP_F64_F32, fpextend>;
183 defm V_CVT_F32_UBYTE0 : VOP1Inst <"v_cvt_f32_ubyte0", VOP1_F32_I32, AMDGPUcvt_f32_ubyte0>;
184 defm V_CVT_F32_UBYTE1 : VOP1Inst <"v_cvt_f32_ubyte1", VOP1_F32_I32, AMDGPUcvt_f32_ubyte1>;
185 defm V_CVT_F32_UBYTE2 : VOP1Inst <"v_cvt_f32_ubyte2", VOP1_F32_I32, AMDGPUcvt_f32_ubyte2>;
186 defm V_CVT_F32_UBYTE3 : VOP1Inst <"v_cvt_f32_ubyte3", VOP1_F32_I32, AMDGPUcvt_f32_ubyte3>;
187 defm V_CVT_U32_F64 : VOP1Inst <"v_cvt_u32_f64", VOP_I32_F64, fp_to_uint>;
188 defm V_CVT_F64_U32 : VOP1Inst <"v_cvt_f64_u32", VOP1_F64_I32, uint_to_fp>;
189 } // End SchedRW = [WriteQuarterRate32]
191 defm V_FRACT_F32 : VOP1Inst <"v_fract_f32", VOP_F32_F32, AMDGPUfract>;
192 defm V_TRUNC_F32 : VOP1Inst <"v_trunc_f32", VOP_F32_F32, ftrunc>;
193 defm V_CEIL_F32 : VOP1Inst <"v_ceil_f32", VOP_F32_F32, fceil>;
194 defm V_RNDNE_F32 : VOP1Inst <"v_rndne_f32", VOP_F32_F32, frint>;
195 defm V_FLOOR_F32 : VOP1Inst <"v_floor_f32", VOP_F32_F32, ffloor>;
197 let SchedRW = [WriteQuarterRate32] in {
198 defm V_EXP_F32 : VOP1Inst <"v_exp_f32", VOP_F32_F32, fexp2>;
199 defm V_LOG_F32 : VOP1Inst <"v_log_f32", VOP_F32_F32, flog2>;
200 defm V_RCP_F32 : VOP1Inst <"v_rcp_f32", VOP_F32_F32, AMDGPUrcp>;
201 defm V_RCP_IFLAG_F32 : VOP1Inst <"v_rcp_iflag_f32", VOP_F32_F32, AMDGPUrcp_iflag>;
202 defm V_RSQ_F32 : VOP1Inst <"v_rsq_f32", VOP_F32_F32, AMDGPUrsq>;
203 defm V_SQRT_F32 : VOP1Inst <"v_sqrt_f32", VOP_F32_F32, fsqrt>;
204 } // End SchedRW = [WriteQuarterRate32]
206 let SchedRW = [WriteDouble] in {
207 defm V_RCP_F64 : VOP1Inst <"v_rcp_f64", VOP_F64_F64, AMDGPUrcp>;
208 defm V_RSQ_F64 : VOP1Inst <"v_rsq_f64", VOP_F64_F64, AMDGPUrsq>;
209 } // End SchedRW = [WriteDouble];
211 let SchedRW = [WriteDouble] in {
212 defm V_SQRT_F64 : VOP1Inst <"v_sqrt_f64", VOP_F64_F64, fsqrt>;
213 } // End SchedRW = [WriteDouble]
215 let SchedRW = [WriteQuarterRate32] in {
216 defm V_SIN_F32 : VOP1Inst <"v_sin_f32", VOP_F32_F32, AMDGPUsin>;
217 defm V_COS_F32 : VOP1Inst <"v_cos_f32", VOP_F32_F32, AMDGPUcos>;
218 } // End SchedRW = [WriteQuarterRate32]
220 defm V_NOT_B32 : VOP1Inst <"v_not_b32", VOP_I32_I32>;
221 defm V_BFREV_B32 : VOP1Inst <"v_bfrev_b32", VOP_I32_I32>;
222 defm V_FFBH_U32 : VOP1Inst <"v_ffbh_u32", VOP_I32_I32>;
223 defm V_FFBL_B32 : VOP1Inst <"v_ffbl_b32", VOP_I32_I32>;
224 defm V_FFBH_I32 : VOP1Inst <"v_ffbh_i32", VOP_I32_I32>;
226 let SchedRW = [WriteDoubleAdd] in {
227 defm V_FREXP_EXP_I32_F64 : VOP1Inst <"v_frexp_exp_i32_f64", VOP_I32_F64, int_amdgcn_frexp_exp>;
228 defm V_FREXP_MANT_F64 : VOP1Inst <"v_frexp_mant_f64", VOP_F64_F64, int_amdgcn_frexp_mant>;
229 defm V_FRACT_F64 : VOP1Inst <"v_fract_f64", VOP_F64_F64, AMDGPUfract>;
230 } // End SchedRW = [WriteDoubleAdd]
232 defm V_FREXP_EXP_I32_F32 : VOP1Inst <"v_frexp_exp_i32_f32", VOP_I32_F32, int_amdgcn_frexp_exp>;
233 defm V_FREXP_MANT_F32 : VOP1Inst <"v_frexp_mant_f32", VOP_F32_F32, int_amdgcn_frexp_mant>;
235 let VOPAsmPrefer32Bit = 1 in {
236 defm V_CLREXCP : VOP1Inst <"v_clrexcp", VOP_NO_EXT<VOP_NONE>>;
239 // Restrict src0 to be VGPR
240 def VOP_I32_VI32_NO_EXT : VOPProfile<[i32, i32, untyped, untyped]> {
241 let Src0RC32 = VRegSrc_32;
242 let Src0RC64 = VRegSrc_32;
248 // Special case because there are no true output operands. Hack vdst
249 // to be a src operand. The custom inserter must add a tied implicit
250 // def and use of the super register since there seems to be no way to
251 // add an implicit def of a virtual register in tablegen.
252 def VOP_MOVRELD : VOPProfile<[untyped, i32, untyped, untyped]> {
253 let Src0RC32 = VOPDstOperand<VGPR_32>;
254 let Src0RC64 = VOPDstOperand<VGPR_32>;
257 let Ins32 = (ins Src0RC32:$vdst, VSrc_b32:$src0);
258 let Ins64 = (ins Src0RC64:$vdst, VSrc_b32:$src0);
259 let InsDPP = (ins DstRC:$vdst, DstRC:$old, Src0RC32:$src0,
260 dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
261 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl);
263 let InsSDWA = (ins Src0RC32:$vdst, Src0ModSDWA:$src0_modifiers, Src0SDWA:$src0,
264 clampmod:$clamp, omod:$omod, dst_sel:$dst_sel, dst_unused:$dst_unused,
267 let Asm32 = getAsm32<1, 1>.ret;
268 let Asm64 = getAsm64<1, 1, 0, 0, 1>.ret;
269 let AsmDPP = getAsmDPP<1, 1, 0>.ret;
270 let AsmSDWA = getAsmSDWA<1, 1>.ret;
271 let AsmSDWA9 = getAsmSDWA9<1, 0, 1>.ret;
276 let EmitDst = 1; // force vdst emission
279 let SubtargetPredicate = HasMovrel, Uses = [M0, EXEC] in {
280 // v_movreld_b32 is a special case because the destination output
281 // register is really a source. It isn't actually read (but may be
282 // written), and is only to provide the base register to start
283 // indexing from. Tablegen seems to not let you define an implicit
284 // virtual register output for the super register being written into,
285 // so this must have an implicit def of the register added to it.
286 defm V_MOVRELD_B32 : VOP1Inst <"v_movreld_b32", VOP_MOVRELD>;
287 defm V_MOVRELS_B32 : VOP1Inst <"v_movrels_b32", VOP_I32_VI32_NO_EXT>;
288 defm V_MOVRELSD_B32 : VOP1Inst <"v_movrelsd_b32", VOP_NO_EXT<VOP_I32_I32>>;
289 } // End Uses = [M0, EXEC]
291 defm V_MOV_FED_B32 : VOP1Inst <"v_mov_fed_b32", VOP_I32_I32>;
293 // These instruction only exist on SI and CI
294 let SubtargetPredicate = isSICI in {
296 let SchedRW = [WriteQuarterRate32] in {
297 defm V_LOG_CLAMP_F32 : VOP1Inst <"v_log_clamp_f32", VOP_F32_F32, int_amdgcn_log_clamp>;
298 defm V_RCP_CLAMP_F32 : VOP1Inst <"v_rcp_clamp_f32", VOP_F32_F32>;
299 defm V_RCP_LEGACY_F32 : VOP1Inst <"v_rcp_legacy_f32", VOP_F32_F32, AMDGPUrcp_legacy>;
300 defm V_RSQ_CLAMP_F32 : VOP1Inst <"v_rsq_clamp_f32", VOP_F32_F32, AMDGPUrsq_clamp>;
301 defm V_RSQ_LEGACY_F32 : VOP1Inst <"v_rsq_legacy_f32", VOP_F32_F32, AMDGPUrsq_legacy>;
302 } // End SchedRW = [WriteQuarterRate32]
304 let SchedRW = [WriteDouble] in {
305 defm V_RCP_CLAMP_F64 : VOP1Inst <"v_rcp_clamp_f64", VOP_F64_F64>;
306 defm V_RSQ_CLAMP_F64 : VOP1Inst <"v_rsq_clamp_f64", VOP_F64_F64, AMDGPUrsq_clamp>;
307 } // End SchedRW = [WriteDouble]
309 } // End SubtargetPredicate = isSICI
312 let SubtargetPredicate = isCIVI in {
314 let SchedRW = [WriteDoubleAdd] in {
315 defm V_TRUNC_F64 : VOP1Inst <"v_trunc_f64", VOP_F64_F64, ftrunc>;
316 defm V_CEIL_F64 : VOP1Inst <"v_ceil_f64", VOP_F64_F64, fceil>;
317 defm V_FLOOR_F64 : VOP1Inst <"v_floor_f64", VOP_F64_F64, ffloor>;
318 defm V_RNDNE_F64 : VOP1Inst <"v_rndne_f64", VOP_F64_F64, frint>;
319 } // End SchedRW = [WriteDoubleAdd]
321 let SchedRW = [WriteQuarterRate32] in {
322 defm V_LOG_LEGACY_F32 : VOP1Inst <"v_log_legacy_f32", VOP_F32_F32>;
323 defm V_EXP_LEGACY_F32 : VOP1Inst <"v_exp_legacy_f32", VOP_F32_F32>;
324 } // End SchedRW = [WriteQuarterRate32]
326 } // End SubtargetPredicate = isCIVI
329 let SubtargetPredicate = Has16BitInsts in {
331 defm V_CVT_F16_U16 : VOP1Inst <"v_cvt_f16_u16", VOP1_F16_I16, uint_to_fp>;
332 defm V_CVT_F16_I16 : VOP1Inst <"v_cvt_f16_i16", VOP1_F16_I16, sint_to_fp>;
333 defm V_CVT_U16_F16 : VOP1Inst <"v_cvt_u16_f16", VOP_I16_F16, fp_to_uint>;
334 defm V_CVT_I16_F16 : VOP1Inst <"v_cvt_i16_f16", VOP_I16_F16, fp_to_sint>;
335 let SchedRW = [WriteQuarterRate32] in {
336 defm V_RCP_F16 : VOP1Inst <"v_rcp_f16", VOP_F16_F16, AMDGPUrcp>;
337 defm V_SQRT_F16 : VOP1Inst <"v_sqrt_f16", VOP_F16_F16, fsqrt>;
338 defm V_RSQ_F16 : VOP1Inst <"v_rsq_f16", VOP_F16_F16, AMDGPUrsq>;
339 defm V_LOG_F16 : VOP1Inst <"v_log_f16", VOP_F16_F16, flog2>;
340 defm V_EXP_F16 : VOP1Inst <"v_exp_f16", VOP_F16_F16, fexp2>;
341 defm V_SIN_F16 : VOP1Inst <"v_sin_f16", VOP_F16_F16, AMDGPUsin>;
342 defm V_COS_F16 : VOP1Inst <"v_cos_f16", VOP_F16_F16, AMDGPUcos>;
343 } // End SchedRW = [WriteQuarterRate32]
344 defm V_FREXP_MANT_F16 : VOP1Inst <"v_frexp_mant_f16", VOP_F16_F16, int_amdgcn_frexp_mant>;
345 defm V_FREXP_EXP_I16_F16 : VOP1Inst <"v_frexp_exp_i16_f16", VOP_I16_F16, int_amdgcn_frexp_exp>;
346 defm V_FLOOR_F16 : VOP1Inst <"v_floor_f16", VOP_F16_F16, ffloor>;
347 defm V_CEIL_F16 : VOP1Inst <"v_ceil_f16", VOP_F16_F16, fceil>;
348 defm V_TRUNC_F16 : VOP1Inst <"v_trunc_f16", VOP_F16_F16, ftrunc>;
349 defm V_RNDNE_F16 : VOP1Inst <"v_rndne_f16", VOP_F16_F16, frint>;
350 defm V_FRACT_F16 : VOP1Inst <"v_fract_f16", VOP_F16_F16, AMDGPUfract>;
354 let OtherPredicates = [Has16BitInsts] in {
357 (f32 (f16_to_fp i16:$src)),
358 (V_CVT_F32_F16_e32 $src)
362 (i16 (AMDGPUfp_to_f16 f32:$src)),
363 (V_CVT_F16_F32_e32 $src)
368 def VOP_SWAP_I32 : VOPProfile<[i32, i32, i32, untyped]> {
369 let Outs32 = (outs VGPR_32:$vdst, VGPR_32:$vdst1);
370 let Ins32 = (ins VGPR_32:$src0, VGPR_32:$src1);
372 let Asm32 = " $vdst, $src0";
377 let SubtargetPredicate = isGFX9 in {
378 let Constraints = "$vdst = $src1, $vdst1 = $src0",
379 DisableEncoding="$vdst1,$src1",
380 SchedRW = [Write64Bit, Write64Bit] in {
381 // Never VOP3. Takes as long as 2 v_mov_b32s
382 def V_SWAP_B32 : VOP1_Pseudo <"v_swap_b32", VOP_SWAP_I32, [], 1>;
385 defm V_SCREEN_PARTITION_4SE_B32 : VOP1Inst <"v_screen_partition_4se_b32", VOP_I32_I32>;
387 defm V_SAT_PK_U8_I16 : VOP1Inst<"v_sat_pk_u8_i16", VOP_I32_I32>;
388 defm V_CVT_NORM_I16_F16 : VOP1Inst<"v_cvt_norm_i16_f16", VOP_I16_F16>;
389 defm V_CVT_NORM_U16_F16 : VOP1Inst<"v_cvt_norm_u16_f16", VOP_I16_F16>;
391 } // End SubtargetPredicate = isGFX9
393 //===----------------------------------------------------------------------===//
395 //===----------------------------------------------------------------------===//
397 //===----------------------------------------------------------------------===//
399 //===----------------------------------------------------------------------===//
401 multiclass VOP1_Real_si <bits<9> op> {
402 let AssemblerPredicates = [isSICI], DecoderNamespace = "SICI" in {
404 VOP1_Real<!cast<VOP1_Pseudo>(NAME#"_e32"), SIEncodingFamily.SI>,
405 VOP1e<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32").Pfl>;
407 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>,
408 VOP3e_si <{1, 1, op{6-0}}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
412 defm V_NOP : VOP1_Real_si <0x0>;
413 defm V_MOV_B32 : VOP1_Real_si <0x1>;
414 defm V_CVT_I32_F64 : VOP1_Real_si <0x3>;
415 defm V_CVT_F64_I32 : VOP1_Real_si <0x4>;
416 defm V_CVT_F32_I32 : VOP1_Real_si <0x5>;
417 defm V_CVT_F32_U32 : VOP1_Real_si <0x6>;
418 defm V_CVT_U32_F32 : VOP1_Real_si <0x7>;
419 defm V_CVT_I32_F32 : VOP1_Real_si <0x8>;
420 defm V_MOV_FED_B32 : VOP1_Real_si <0x9>;
421 defm V_CVT_F16_F32 : VOP1_Real_si <0xa>;
422 defm V_CVT_F32_F16 : VOP1_Real_si <0xb>;
423 defm V_CVT_RPI_I32_F32 : VOP1_Real_si <0xc>;
424 defm V_CVT_FLR_I32_F32 : VOP1_Real_si <0xd>;
425 defm V_CVT_OFF_F32_I4 : VOP1_Real_si <0xe>;
426 defm V_CVT_F32_F64 : VOP1_Real_si <0xf>;
427 defm V_CVT_F64_F32 : VOP1_Real_si <0x10>;
428 defm V_CVT_F32_UBYTE0 : VOP1_Real_si <0x11>;
429 defm V_CVT_F32_UBYTE1 : VOP1_Real_si <0x12>;
430 defm V_CVT_F32_UBYTE2 : VOP1_Real_si <0x13>;
431 defm V_CVT_F32_UBYTE3 : VOP1_Real_si <0x14>;
432 defm V_CVT_U32_F64 : VOP1_Real_si <0x15>;
433 defm V_CVT_F64_U32 : VOP1_Real_si <0x16>;
434 defm V_FRACT_F32 : VOP1_Real_si <0x20>;
435 defm V_TRUNC_F32 : VOP1_Real_si <0x21>;
436 defm V_CEIL_F32 : VOP1_Real_si <0x22>;
437 defm V_RNDNE_F32 : VOP1_Real_si <0x23>;
438 defm V_FLOOR_F32 : VOP1_Real_si <0x24>;
439 defm V_EXP_F32 : VOP1_Real_si <0x25>;
440 defm V_LOG_CLAMP_F32 : VOP1_Real_si <0x26>;
441 defm V_LOG_F32 : VOP1_Real_si <0x27>;
442 defm V_RCP_CLAMP_F32 : VOP1_Real_si <0x28>;
443 defm V_RCP_LEGACY_F32 : VOP1_Real_si <0x29>;
444 defm V_RCP_F32 : VOP1_Real_si <0x2a>;
445 defm V_RCP_IFLAG_F32 : VOP1_Real_si <0x2b>;
446 defm V_RSQ_CLAMP_F32 : VOP1_Real_si <0x2c>;
447 defm V_RSQ_LEGACY_F32 : VOP1_Real_si <0x2d>;
448 defm V_RSQ_F32 : VOP1_Real_si <0x2e>;
449 defm V_RCP_F64 : VOP1_Real_si <0x2f>;
450 defm V_RCP_CLAMP_F64 : VOP1_Real_si <0x30>;
451 defm V_RSQ_F64 : VOP1_Real_si <0x31>;
452 defm V_RSQ_CLAMP_F64 : VOP1_Real_si <0x32>;
453 defm V_SQRT_F32 : VOP1_Real_si <0x33>;
454 defm V_SQRT_F64 : VOP1_Real_si <0x34>;
455 defm V_SIN_F32 : VOP1_Real_si <0x35>;
456 defm V_COS_F32 : VOP1_Real_si <0x36>;
457 defm V_NOT_B32 : VOP1_Real_si <0x37>;
458 defm V_BFREV_B32 : VOP1_Real_si <0x38>;
459 defm V_FFBH_U32 : VOP1_Real_si <0x39>;
460 defm V_FFBL_B32 : VOP1_Real_si <0x3a>;
461 defm V_FFBH_I32 : VOP1_Real_si <0x3b>;
462 defm V_FREXP_EXP_I32_F64 : VOP1_Real_si <0x3c>;
463 defm V_FREXP_MANT_F64 : VOP1_Real_si <0x3d>;
464 defm V_FRACT_F64 : VOP1_Real_si <0x3e>;
465 defm V_FREXP_EXP_I32_F32 : VOP1_Real_si <0x3f>;
466 defm V_FREXP_MANT_F32 : VOP1_Real_si <0x40>;
467 defm V_CLREXCP : VOP1_Real_si <0x41>;
468 defm V_MOVRELD_B32 : VOP1_Real_si <0x42>;
469 defm V_MOVRELS_B32 : VOP1_Real_si <0x43>;
470 defm V_MOVRELSD_B32 : VOP1_Real_si <0x44>;
472 //===----------------------------------------------------------------------===//
474 //===----------------------------------------------------------------------===//
476 multiclass VOP1_Real_ci <bits<9> op> {
477 let AssemblerPredicates = [isCIOnly], DecoderNamespace = "CI" in {
479 VOP1_Real<!cast<VOP1_Pseudo>(NAME#"_e32"), SIEncodingFamily.SI>,
480 VOP1e<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32").Pfl>;
482 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>,
483 VOP3e_si <{1, 1, op{6-0}}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
487 defm V_TRUNC_F64 : VOP1_Real_ci <0x17>;
488 defm V_CEIL_F64 : VOP1_Real_ci <0x18>;
489 defm V_FLOOR_F64 : VOP1_Real_ci <0x1A>;
490 defm V_RNDNE_F64 : VOP1_Real_ci <0x19>;
491 defm V_LOG_LEGACY_F32 : VOP1_Real_ci <0x45>;
492 defm V_EXP_LEGACY_F32 : VOP1_Real_ci <0x46>;
494 //===----------------------------------------------------------------------===//
496 //===----------------------------------------------------------------------===//
498 class VOP1_DPP <bits<8> op, VOP1_Pseudo ps, VOPProfile P = ps.Pfl> :
499 VOP_DPP <ps.OpName, P> {
502 let SchedRW = ps.SchedRW;
503 let hasSideEffects = ps.hasSideEffects;
506 let Inst{8-0} = 0xfa; // dpp
508 let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
509 let Inst{31-25} = 0x3f; //encoding
512 multiclass VOP1Only_Real_vi <bits<10> op> {
513 let AssemblerPredicates = [isVI], DecoderNamespace = "VI" in {
515 VOP1_Real<!cast<VOP1_Pseudo>(NAME), SIEncodingFamily.VI>,
516 VOP1e<op{7-0}, !cast<VOP1_Pseudo>(NAME).Pfl>;
520 multiclass VOP1_Real_e32e64_vi <bits<10> op> {
521 let AssemblerPredicates = [isVI], DecoderNamespace = "VI" in {
523 VOP1_Real<!cast<VOP1_Pseudo>(NAME#"_e32"), SIEncodingFamily.VI>,
524 VOP1e<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32").Pfl>;
526 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
527 VOP3e_vi <!add(0x140, op), !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
531 multiclass VOP1_Real_vi <bits<10> op> {
532 defm NAME : VOP1_Real_e32e64_vi <op>;
535 VOP_SDWA_Real <!cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa")>,
536 VOP1_SDWAe <op{7-0}, !cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa").Pfl>;
539 VOP_SDWA9_Real <!cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa")>,
540 VOP1_SDWA9Ae <op{7-0}, !cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa").Pfl>;
542 // For now left dpp only for asm/dasm
543 // TODO: add corresponding pseudo
544 def _dpp : VOP1_DPP<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32")>;
547 defm V_NOP : VOP1_Real_vi <0x0>;
548 defm V_MOV_B32 : VOP1_Real_vi <0x1>;
549 defm V_CVT_I32_F64 : VOP1_Real_vi <0x3>;
550 defm V_CVT_F64_I32 : VOP1_Real_vi <0x4>;
551 defm V_CVT_F32_I32 : VOP1_Real_vi <0x5>;
552 defm V_CVT_F32_U32 : VOP1_Real_vi <0x6>;
553 defm V_CVT_U32_F32 : VOP1_Real_vi <0x7>;
554 defm V_CVT_I32_F32 : VOP1_Real_vi <0x8>;
555 defm V_MOV_FED_B32 : VOP1_Real_vi <0x9>;
556 defm V_CVT_F16_F32 : VOP1_Real_vi <0xa>;
557 defm V_CVT_F32_F16 : VOP1_Real_vi <0xb>;
558 defm V_CVT_RPI_I32_F32 : VOP1_Real_vi <0xc>;
559 defm V_CVT_FLR_I32_F32 : VOP1_Real_vi <0xd>;
560 defm V_CVT_OFF_F32_I4 : VOP1_Real_vi <0xe>;
561 defm V_CVT_F32_F64 : VOP1_Real_vi <0xf>;
562 defm V_CVT_F64_F32 : VOP1_Real_vi <0x10>;
563 defm V_CVT_F32_UBYTE0 : VOP1_Real_vi <0x11>;
564 defm V_CVT_F32_UBYTE1 : VOP1_Real_vi <0x12>;
565 defm V_CVT_F32_UBYTE2 : VOP1_Real_vi <0x13>;
566 defm V_CVT_F32_UBYTE3 : VOP1_Real_vi <0x14>;
567 defm V_CVT_U32_F64 : VOP1_Real_vi <0x15>;
568 defm V_CVT_F64_U32 : VOP1_Real_vi <0x16>;
569 defm V_FRACT_F32 : VOP1_Real_vi <0x1b>;
570 defm V_TRUNC_F32 : VOP1_Real_vi <0x1c>;
571 defm V_CEIL_F32 : VOP1_Real_vi <0x1d>;
572 defm V_RNDNE_F32 : VOP1_Real_vi <0x1e>;
573 defm V_FLOOR_F32 : VOP1_Real_vi <0x1f>;
574 defm V_EXP_F32 : VOP1_Real_vi <0x20>;
575 defm V_LOG_F32 : VOP1_Real_vi <0x21>;
576 defm V_RCP_F32 : VOP1_Real_vi <0x22>;
577 defm V_RCP_IFLAG_F32 : VOP1_Real_vi <0x23>;
578 defm V_RSQ_F32 : VOP1_Real_vi <0x24>;
579 defm V_RCP_F64 : VOP1_Real_vi <0x25>;
580 defm V_RSQ_F64 : VOP1_Real_vi <0x26>;
581 defm V_SQRT_F32 : VOP1_Real_vi <0x27>;
582 defm V_SQRT_F64 : VOP1_Real_vi <0x28>;
583 defm V_SIN_F32 : VOP1_Real_vi <0x29>;
584 defm V_COS_F32 : VOP1_Real_vi <0x2a>;
585 defm V_NOT_B32 : VOP1_Real_vi <0x2b>;
586 defm V_BFREV_B32 : VOP1_Real_vi <0x2c>;
587 defm V_FFBH_U32 : VOP1_Real_vi <0x2d>;
588 defm V_FFBL_B32 : VOP1_Real_vi <0x2e>;
589 defm V_FFBH_I32 : VOP1_Real_vi <0x2f>;
590 defm V_FREXP_EXP_I32_F64 : VOP1_Real_vi <0x30>;
591 defm V_FREXP_MANT_F64 : VOP1_Real_vi <0x31>;
592 defm V_FRACT_F64 : VOP1_Real_vi <0x32>;
593 defm V_FREXP_EXP_I32_F32 : VOP1_Real_vi <0x33>;
594 defm V_FREXP_MANT_F32 : VOP1_Real_vi <0x34>;
595 defm V_CLREXCP : VOP1_Real_vi <0x35>;
596 defm V_MOVRELD_B32 : VOP1_Real_e32e64_vi <0x36>;
597 defm V_MOVRELS_B32 : VOP1_Real_e32e64_vi <0x37>;
598 defm V_MOVRELSD_B32 : VOP1_Real_e32e64_vi <0x38>;
599 defm V_TRUNC_F64 : VOP1_Real_vi <0x17>;
600 defm V_CEIL_F64 : VOP1_Real_vi <0x18>;
601 defm V_FLOOR_F64 : VOP1_Real_vi <0x1A>;
602 defm V_RNDNE_F64 : VOP1_Real_vi <0x19>;
603 defm V_LOG_LEGACY_F32 : VOP1_Real_vi <0x4c>;
604 defm V_EXP_LEGACY_F32 : VOP1_Real_vi <0x4b>;
605 defm V_CVT_F16_U16 : VOP1_Real_vi <0x39>;
606 defm V_CVT_F16_I16 : VOP1_Real_vi <0x3a>;
607 defm V_CVT_U16_F16 : VOP1_Real_vi <0x3b>;
608 defm V_CVT_I16_F16 : VOP1_Real_vi <0x3c>;
609 defm V_RCP_F16 : VOP1_Real_vi <0x3d>;
610 defm V_SQRT_F16 : VOP1_Real_vi <0x3e>;
611 defm V_RSQ_F16 : VOP1_Real_vi <0x3f>;
612 defm V_LOG_F16 : VOP1_Real_vi <0x40>;
613 defm V_EXP_F16 : VOP1_Real_vi <0x41>;
614 defm V_FREXP_MANT_F16 : VOP1_Real_vi <0x42>;
615 defm V_FREXP_EXP_I16_F16 : VOP1_Real_vi <0x43>;
616 defm V_FLOOR_F16 : VOP1_Real_vi <0x44>;
617 defm V_CEIL_F16 : VOP1_Real_vi <0x45>;
618 defm V_TRUNC_F16 : VOP1_Real_vi <0x46>;
619 defm V_RNDNE_F16 : VOP1_Real_vi <0x47>;
620 defm V_FRACT_F16 : VOP1_Real_vi <0x48>;
621 defm V_SIN_F16 : VOP1_Real_vi <0x49>;
622 defm V_COS_F16 : VOP1_Real_vi <0x4a>;
623 defm V_SWAP_B32 : VOP1Only_Real_vi <0x51>;
625 defm V_SAT_PK_U8_I16 : VOP1_Real_vi<0x4f>;
626 defm V_CVT_NORM_I16_F16 : VOP1_Real_vi<0x4d>;
627 defm V_CVT_NORM_U16_F16 : VOP1_Real_vi<0x4e>;
629 // Copy of v_mov_b32 with $vdst as a use operand for use with VGPR
630 // indexing mode. vdst can't be treated as a def for codegen purposes,
631 // and an implicit use and def of the super register should be added.
632 def V_MOV_B32_indirect : VPseudoInstSI<(outs),
633 (ins getVALUDstForVT<i32>.ret:$vdst, getVOPSrc0ForVT<i32>.ret:$src0)>,
634 PseudoInstExpansion<(V_MOV_B32_e32_vi getVALUDstForVT<i32>.ret:$vdst,
635 getVOPSrc0ForVT<i32>.ret:$src0)> {
637 let SubtargetPredicate = isVI;
640 // This is a pseudo variant of the v_movreld_b32 instruction in which the
641 // vector operand appears only twice, once as def and once as use. Using this
642 // pseudo avoids problems with the Two Address instructions pass.
643 class V_MOVRELD_B32_pseudo<RegisterClass rc> : VPseudoInstSI <
645 (ins rc:$vsrc, VSrc_b32:$val, i32imm:$offset)> {
648 let Constraints = "$vsrc = $vdst";
649 let Uses = [M0, EXEC];
651 let SubtargetPredicate = HasMovrel;
654 def V_MOVRELD_B32_V1 : V_MOVRELD_B32_pseudo<VGPR_32>;
655 def V_MOVRELD_B32_V2 : V_MOVRELD_B32_pseudo<VReg_64>;
656 def V_MOVRELD_B32_V4 : V_MOVRELD_B32_pseudo<VReg_128>;
657 def V_MOVRELD_B32_V8 : V_MOVRELD_B32_pseudo<VReg_256>;
658 def V_MOVRELD_B32_V16 : V_MOVRELD_B32_pseudo<VReg_512>;
660 let OtherPredicates = [isVI] in {
663 (i32 (int_amdgcn_mov_dpp i32:$src, imm:$dpp_ctrl, imm:$row_mask, imm:$bank_mask,
665 (V_MOV_B32_dpp $src, $src, (as_i32imm $dpp_ctrl),
666 (as_i32imm $row_mask), (as_i32imm $bank_mask),
667 (as_i1imm $bound_ctrl))
671 (i32 (int_amdgcn_update_dpp i32:$old, i32:$src, imm:$dpp_ctrl, imm:$row_mask,
672 imm:$bank_mask, imm:$bound_ctrl)),
673 (V_MOV_B32_dpp $old, $src, (as_i32imm $dpp_ctrl),
674 (as_i32imm $row_mask), (as_i32imm $bank_mask),
675 (as_i1imm $bound_ctrl))
679 (i32 (anyext i16:$src)),
684 (i64 (anyext i16:$src)),
685 (REG_SEQUENCE VReg_64,
686 (i32 (COPY $src)), sub0,
687 (V_MOV_B32_e32 (i32 0)), sub1)
691 (i16 (trunc i32:$src)),
696 (i16 (trunc i64:$src)),
697 (EXTRACT_SUBREG $src, sub0)
700 } // End OtherPredicates = [isVI]
702 //===----------------------------------------------------------------------===//
704 //===----------------------------------------------------------------------===//
706 multiclass VOP1_Real_gfx9 <bits<10> op> {
707 let AssemblerPredicates = [isGFX9], DecoderNamespace = "GFX9" in {
708 defm NAME : VOP1_Real_e32e64_vi <op>;
712 VOP_SDWA9_Real <!cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa")>,
713 VOP1_SDWA9Ae <op{7-0}, !cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa").Pfl>;
715 // For now left dpp only for asm/dasm
716 // TODO: add corresponding pseudo
717 def _dpp : VOP1_DPP<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32")>;
720 defm V_SCREEN_PARTITION_4SE_B32 : VOP1_Real_gfx9 <0x37>;