1 //===-- VOP3Instructions.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 getVOP3ModPat<VOPProfile P, SDPatternOperator node> {
15 dag src0 = !if(P.HasOMod,
16 (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp, i32:$omod),
17 (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp));
19 list<dag> ret3 = [(set P.DstVT:$vdst,
20 (DivergentFragOrOp<node, P>.ret (P.Src0VT src0),
21 (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers)),
22 (P.Src2VT (VOP3Mods P.Src2VT:$src2, i32:$src2_modifiers))))];
24 list<dag> ret2 = [(set P.DstVT:$vdst,
25 (DivergentFragOrOp<node, P>.ret (P.Src0VT src0),
26 (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers))))];
28 list<dag> ret1 = [(set P.DstVT:$vdst,
29 (DivergentFragOrOp<node, P>.ret (P.Src0VT src0)))];
31 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3,
32 !if(!eq(P.NumSrcArgs, 2), ret2,
36 class getVOP3PModPat<VOPProfile P, SDPatternOperator node> {
37 list<dag> ret3 = [(set P.DstVT:$vdst,
38 (DivergentFragOrOp<node, P>.ret (P.Src0VT !if(P.HasClamp, (VOP3PMods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp),
39 (VOP3PMods P.Src0VT:$src0, i32:$src0_modifiers))),
40 (P.Src1VT (VOP3PMods P.Src1VT:$src1, i32:$src1_modifiers)),
41 (P.Src2VT (VOP3PMods P.Src2VT:$src2, i32:$src2_modifiers))))];
43 list<dag> ret2 = [(set P.DstVT:$vdst,
44 (DivergentFragOrOp<node, P>.ret !if(P.HasClamp, (P.Src0VT (VOP3PMods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp)),
45 (P.Src0VT (VOP3PMods P.Src0VT:$src0, i32:$src0_modifiers))),
46 (P.Src1VT (VOP3PMods P.Src1VT:$src1, i32:$src1_modifiers))))];
48 list<dag> ret1 = [(set P.DstVT:$vdst,
49 (DivergentFragOrOp<node, P>.ret (P.Src0VT (VOP3PMods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp))))];
51 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3,
52 !if(!eq(P.NumSrcArgs, 2), ret2,
56 class getVOP3OpSelPat<VOPProfile P, SDPatternOperator node> {
57 list<dag> ret3 = [(set P.DstVT:$vdst,
58 (DivergentFragOrOp<node, P>.ret (P.Src0VT !if(P.HasClamp, (VOP3OpSel0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp),
59 (VOP3OpSel P.Src0VT:$src0, i32:$src0_modifiers))),
60 (P.Src1VT (VOP3OpSel P.Src1VT:$src1, i32:$src1_modifiers)),
61 (P.Src2VT (VOP3OpSel P.Src2VT:$src2, i32:$src2_modifiers))))];
63 list<dag> ret2 = [(set P.DstVT:$vdst,
64 (DivergentFragOrOp<node, P>.ret !if(P.HasClamp, (P.Src0VT (VOP3OpSel0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp)),
65 (P.Src0VT (VOP3OpSel P.Src0VT:$src0, i32:$src0_modifiers))),
66 (P.Src1VT (VOP3OpSel P.Src1VT:$src1, i32:$src1_modifiers))))];
68 list<dag> ret1 = [(set P.DstVT:$vdst,
69 (DivergentFragOrOp<node, P>.ret (P.Src0VT (VOP3OpSel0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp))))];
71 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3,
72 !if(!eq(P.NumSrcArgs, 2), ret2,
76 class getVOP3OpSelModPat<VOPProfile P, SDPatternOperator node> {
77 list<dag> ret3 = [(set P.DstVT:$vdst,
78 (DivergentFragOrOp<node, P>.ret (P.Src0VT !if(P.HasClamp, (VOP3OpSelMods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp),
79 (VOP3OpSelMods P.Src0VT:$src0, i32:$src0_modifiers))),
80 (P.Src1VT (VOP3OpSelMods P.Src1VT:$src1, i32:$src1_modifiers)),
81 (P.Src2VT (VOP3OpSelMods P.Src2VT:$src2, i32:$src2_modifiers))))];
83 list<dag> ret2 = [(set P.DstVT:$vdst,
84 (DivergentFragOrOp<node, P>.ret !if(P.HasClamp, (P.Src0VT (VOP3OpSelMods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp)),
85 (P.Src0VT (VOP3OpSelMods P.Src0VT:$src0, i32:$src0_modifiers))),
86 (P.Src1VT (VOP3OpSelMods P.Src1VT:$src1, i32:$src1_modifiers))))];
88 list<dag> ret1 = [(set P.DstVT:$vdst,
89 (DivergentFragOrOp<node, P>.ret (P.Src0VT (VOP3OpSelMods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp))))];
91 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3,
92 !if(!eq(P.NumSrcArgs, 2), ret2,
96 class getVOP3Pat<VOPProfile P, SDPatternOperator node> {
97 list<dag> ret3 = [(set P.DstVT:$vdst, (DivergentFragOrOp<node, P>.ret P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2))];
98 list<dag> ret2 = [(set P.DstVT:$vdst, (DivergentFragOrOp<node, P>.ret P.Src0VT:$src0, P.Src1VT:$src1))];
99 list<dag> ret1 = [(set P.DstVT:$vdst, (DivergentFragOrOp<node, P>.ret P.Src0VT:$src0))];
100 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3,
101 !if(!eq(P.NumSrcArgs, 2), ret2,
105 class getVOP3ClampPat<VOPProfile P, SDPatternOperator node> {
106 list<dag> ret3 = [(set P.DstVT:$vdst, (node P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2, i1:$clamp))];
107 list<dag> ret2 = [(set P.DstVT:$vdst, (node P.Src0VT:$src0, P.Src1VT:$src1, i1:$clamp))];
108 list<dag> ret1 = [(set P.DstVT:$vdst, (node P.Src0VT:$src0, i1:$clamp))];
109 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3,
110 !if(!eq(P.NumSrcArgs, 2), ret2,
114 class VOP3Inst<string OpName, VOPProfile P, SDPatternOperator node = null_frag, bit VOP3Only = 0> :
115 VOP3_Pseudo<OpName, P,
118 getVOP3OpSelModPat<P, node>.ret,
119 getVOP3OpSelPat<P, node>.ret),
121 getVOP3ModPat<P, node>.ret,
123 getVOP3ClampPat<P, node>.ret,
124 getVOP3Pat<P, node>.ret))),
125 VOP3Only, 0, P.HasOpSel> {
127 let IntClamp = P.HasIntClamp;
128 let AsmMatchConverter =
131 !if(!or(P.HasModifiers, !or(P.HasOMod, P.HasIntClamp)),
136 // Special case for v_div_fmas_{f32|f64}, since it seems to be the
137 // only VOP instruction that implicitly reads VCC.
138 let Asm64 = " $vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$clamp$omod" in {
139 def VOP_F32_F32_F32_F32_VCC : VOPProfile<[f32, f32, f32, f32]> {
140 let Outs64 = (outs DstRC.RegClass:$vdst);
142 def VOP_F64_F64_F64_F64_VCC : VOPProfile<[f64, f64, f64, f64]> {
143 let Outs64 = (outs DstRC.RegClass:$vdst);
147 class getVOP3VCC<VOPProfile P, SDPatternOperator node> {
150 (node (P.Src0VT (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp, i32:$omod)),
151 (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers)),
152 (P.Src2VT (VOP3Mods P.Src2VT:$src2, i32:$src2_modifiers)),
156 class VOP3Features<bit Clamp, bit OpSel, bit Packed> {
157 bit HasClamp = Clamp;
158 bit HasOpSel = OpSel;
159 bit IsPacked = Packed;
162 def VOP3_REGULAR : VOP3Features<0, 0, 0>;
163 def VOP3_CLAMP : VOP3Features<1, 0, 0>;
164 def VOP3_OPSEL : VOP3Features<1, 1, 0>;
165 def VOP3_PACKED : VOP3Features<1, 1, 1>;
167 class VOP3_Profile<VOPProfile P, VOP3Features Features = VOP3_REGULAR> : VOPProfile<P.ArgVT> {
169 let HasClamp = !if(Features.HasClamp, 1, P.HasClamp);
170 let HasOpSel = !if(Features.HasOpSel, 1, P.HasOpSel);
171 let IsPacked = !if(Features.IsPacked, 1, P.IsPacked);
173 let HasModifiers = !if(Features.IsPacked, 1, P.HasModifiers);
175 // FIXME: Hack to stop printing _e64
176 let Outs64 = (outs DstRC.RegClass:$vdst);
178 " " # !if(Features.HasOpSel,
179 getAsmVOP3OpSel<NumSrcArgs,
183 HasSrc2FloatMods>.ret,
184 !if(Features.HasClamp,
185 getAsm64<HasDst, NumSrcArgs, HasIntClamp,
186 HasModifiers, HasOMod, DstVT>.ret,
188 let NeedPatGen = P.NeedPatGen;
191 class VOP3b_Profile<ValueType vt> : VOPProfile<[vt, vt, vt, vt]> {
192 // v_div_scale_{f32|f64} do not support input modifiers.
193 let HasModifiers = 0;
195 let Outs64 = (outs DstRC:$vdst, SReg_64:$sdst);
196 let Asm64 = " $vdst, $sdst, $src0, $src1, $src2";
199 def VOP3b_F32_I1_F32_F32_F32 : VOP3b_Profile<f32> {
200 // FIXME: Hack to stop printing _e64
201 let DstRC = RegisterOperand<VGPR_32>;
204 def VOP3b_F64_I1_F64_F64_F64 : VOP3b_Profile<f64> {
205 // FIXME: Hack to stop printing _e64
206 let DstRC = RegisterOperand<VReg_64>;
209 def VOP3b_I64_I1_I32_I32_I64 : VOPProfile<[i64, i32, i32, i64]> {
212 // FIXME: Hack to stop printing _e64
213 let DstRC = RegisterOperand<VReg_64>;
215 let Outs64 = (outs DstRC:$vdst, SReg_64:$sdst);
216 let Asm64 = " $vdst, $sdst, $src0, $src1, $src2$clamp";
219 //===----------------------------------------------------------------------===//
221 //===----------------------------------------------------------------------===//
223 class VOP3Interp<string OpName, VOPProfile P, list<dag> pattern = []> :
224 VOP3_Pseudo<OpName, P, pattern> {
225 let AsmMatchConverter = "cvtVOP3Interp";
228 def VOP3_INTERP : VOPProfile<[f32, f32, i32, untyped]> {
229 let Ins64 = (ins Src0Mod:$src0_modifiers, VRegSrc_32:$src0,
230 Attr:$attr, AttrChan:$attrchan,
231 clampmod:$clamp, omod:$omod);
233 let Asm64 = "$vdst, $src0_modifiers, $attr$attrchan$clamp$omod";
236 def VOP3_INTERP_MOV : VOPProfile<[f32, i32, i32, untyped]> {
237 let Ins64 = (ins InterpSlot:$src0,
238 Attr:$attr, AttrChan:$attrchan,
239 clampmod:$clamp, omod:$omod);
241 let Asm64 = "$vdst, $src0, $attr$attrchan$clamp$omod";
246 class getInterp16Asm <bit HasSrc2, bit HasOMod> {
247 string src2 = !if(HasSrc2, ", $src2_modifiers", "");
248 string omod = !if(HasOMod, "$omod", "");
250 " $vdst, $src0_modifiers, $attr$attrchan"#src2#"$high$clamp"#omod;
253 class getInterp16Ins <bit HasSrc2, bit HasOMod,
254 Operand Src0Mod, Operand Src2Mod> {
255 dag ret = !if(HasSrc2,
257 (ins Src0Mod:$src0_modifiers, VRegSrc_32:$src0,
258 Attr:$attr, AttrChan:$attrchan,
259 Src2Mod:$src2_modifiers, VRegSrc_32:$src2,
260 highmod:$high, clampmod:$clamp, omod:$omod),
261 (ins Src0Mod:$src0_modifiers, VRegSrc_32:$src0,
262 Attr:$attr, AttrChan:$attrchan,
263 Src2Mod:$src2_modifiers, VRegSrc_32:$src2,
264 highmod:$high, clampmod:$clamp)
266 (ins Src0Mod:$src0_modifiers, VRegSrc_32:$src0,
267 Attr:$attr, AttrChan:$attrchan,
268 highmod:$high, clampmod:$clamp, omod:$omod)
272 class VOP3_INTERP16 <list<ValueType> ArgVT> : VOPProfile<ArgVT> {
274 let HasOMod = !if(!eq(DstVT.Value, f16.Value), 0, 1);
277 let Outs64 = (outs VGPR_32:$vdst);
278 let Ins64 = getInterp16Ins<HasSrc2, HasOMod, Src0Mod, Src2Mod>.ret;
279 let Asm64 = getInterp16Asm<HasSrc2, HasOMod>.ret;
282 //===----------------------------------------------------------------------===//
284 //===----------------------------------------------------------------------===//
286 let isCommutable = 1 in {
288 def V_MAD_LEGACY_F32 : VOP3Inst <"v_mad_legacy_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
289 def V_MAD_F32 : VOP3Inst <"v_mad_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, fmad>;
290 def V_MAD_I32_I24 : VOP3Inst <"v_mad_i32_i24", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
291 def V_MAD_U32_U24 : VOP3Inst <"v_mad_u32_u24", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
292 def V_FMA_F32 : VOP3Inst <"v_fma_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, fma>;
293 def V_LERP_U8 : VOP3Inst <"v_lerp_u8", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_lerp>;
295 let SchedRW = [WriteDoubleAdd] in {
296 let FPDPRounding = 1 in {
297 def V_FMA_F64 : VOP3Inst <"v_fma_f64", VOP3_Profile<VOP_F64_F64_F64_F64>, fma>;
298 def V_ADD_F64 : VOP3Inst <"v_add_f64", VOP3_Profile<VOP_F64_F64_F64>, fadd, 1>;
299 def V_MUL_F64 : VOP3Inst <"v_mul_f64", VOP3_Profile<VOP_F64_F64_F64>, fmul, 1>;
300 } // End FPDPRounding = 1
301 def V_MIN_F64 : VOP3Inst <"v_min_f64", VOP3_Profile<VOP_F64_F64_F64>, fminnum_like, 1>;
302 def V_MAX_F64 : VOP3Inst <"v_max_f64", VOP3_Profile<VOP_F64_F64_F64>, fmaxnum_like, 1>;
303 } // End SchedRW = [WriteDoubleAdd]
305 let SchedRW = [WriteQuarterRate32] in {
306 def V_MUL_LO_U32 : VOP3Inst <"v_mul_lo_u32", VOP3_Profile<VOP_I32_I32_I32>>;
307 def V_MUL_HI_U32 : VOP3Inst <"v_mul_hi_u32", VOP3_Profile<VOP_I32_I32_I32>, mulhu>;
308 def V_MUL_LO_I32 : VOP3Inst <"v_mul_lo_i32", VOP3_Profile<VOP_I32_I32_I32>>;
309 def V_MUL_HI_I32 : VOP3Inst <"v_mul_hi_i32", VOP3_Profile<VOP_I32_I32_I32>, mulhs>;
310 } // End SchedRW = [WriteQuarterRate32]
312 let Uses = [VCC, EXEC] in {
314 // result = src0 * src1 + src2
318 def V_DIV_FMAS_F32 : VOP3_Pseudo <"v_div_fmas_f32", VOP_F32_F32_F32_F32_VCC,
319 getVOP3VCC<VOP_F32_F32_F32_F32_VCC, AMDGPUdiv_fmas>.ret> {
320 let SchedRW = [WriteFloatFMA];
323 // result = src0 * src1 + src2
327 def V_DIV_FMAS_F64 : VOP3_Pseudo <"v_div_fmas_f64", VOP_F64_F64_F64_F64_VCC,
328 getVOP3VCC<VOP_F64_F64_F64_F64_VCC, AMDGPUdiv_fmas>.ret> {
329 let SchedRW = [WriteDouble];
330 let FPDPRounding = 1;
332 } // End Uses = [VCC, EXEC]
334 } // End isCommutable = 1
336 def V_CUBEID_F32 : VOP3Inst <"v_cubeid_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubeid>;
337 def V_CUBESC_F32 : VOP3Inst <"v_cubesc_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubesc>;
338 def V_CUBETC_F32 : VOP3Inst <"v_cubetc_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubetc>;
339 def V_CUBEMA_F32 : VOP3Inst <"v_cubema_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubema>;
340 def V_BFE_U32 : VOP3Inst <"v_bfe_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUbfe_u32>;
341 def V_BFE_I32 : VOP3Inst <"v_bfe_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUbfe_i32>;
342 def V_BFI_B32 : VOP3Inst <"v_bfi_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUbfi>;
343 def V_ALIGNBIT_B32 : VOP3Inst <"v_alignbit_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_alignbit>;
344 def V_ALIGNBYTE_B32 : VOP3Inst <"v_alignbyte_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_alignbyte>;
345 def V_MIN3_F32 : VOP3Inst <"v_min3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmin3>;
346 def V_MIN3_I32 : VOP3Inst <"v_min3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmin3>;
347 def V_MIN3_U32 : VOP3Inst <"v_min3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumin3>;
348 def V_MAX3_F32 : VOP3Inst <"v_max3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmax3>;
349 def V_MAX3_I32 : VOP3Inst <"v_max3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmax3>;
350 def V_MAX3_U32 : VOP3Inst <"v_max3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumax3>;
351 def V_MED3_F32 : VOP3Inst <"v_med3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmed3>;
352 def V_MED3_I32 : VOP3Inst <"v_med3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmed3>;
353 def V_MED3_U32 : VOP3Inst <"v_med3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumed3>;
354 def V_SAD_U8 : VOP3Inst <"v_sad_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
355 def V_SAD_HI_U8 : VOP3Inst <"v_sad_hi_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
356 def V_SAD_U16 : VOP3Inst <"v_sad_u16", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
357 def V_SAD_U32 : VOP3Inst <"v_sad_u32", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
358 def V_CVT_PK_U8_F32 : VOP3Inst<"v_cvt_pk_u8_f32", VOP3_Profile<VOP_I32_F32_I32_I32>, int_amdgcn_cvt_pk_u8_f32>;
359 def V_DIV_FIXUP_F32 : VOP3Inst <"v_div_fixup_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUdiv_fixup>;
361 let SchedRW = [WriteDoubleAdd], FPDPRounding = 1 in {
362 def V_DIV_FIXUP_F64 : VOP3Inst <"v_div_fixup_f64", VOP3_Profile<VOP_F64_F64_F64_F64>, AMDGPUdiv_fixup>;
363 def V_LDEXP_F64 : VOP3Inst <"v_ldexp_f64", VOP3_Profile<VOP_F64_F64_I32>, AMDGPUldexp, 1>;
364 } // End SchedRW = [WriteDoubleAdd], FPDPRounding = 1
366 def V_DIV_SCALE_F32 : VOP3_Pseudo <"v_div_scale_f32", VOP3b_F32_I1_F32_F32_F32, [], 1> {
367 let SchedRW = [WriteFloatFMA, WriteSALU];
368 let AsmMatchConverter = "";
371 // Double precision division pre-scale.
372 def V_DIV_SCALE_F64 : VOP3_Pseudo <"v_div_scale_f64", VOP3b_F64_I1_F64_F64_F64, [], 1> {
373 let SchedRW = [WriteDouble, WriteSALU];
374 let AsmMatchConverter = "";
375 let FPDPRounding = 1;
378 def V_MSAD_U8 : VOP3Inst <"v_msad_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
380 let Constraints = "@earlyclobber $vdst" in {
381 def V_MQSAD_PK_U16_U8 : VOP3Inst <"v_mqsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64, VOP3_CLAMP>>;
382 } // End Constraints = "@earlyclobber $vdst"
384 def V_TRIG_PREOP_F64 : VOP3Inst <"v_trig_preop_f64", VOP3_Profile<VOP_F64_F64_I32>, AMDGPUtrig_preop> {
385 let SchedRW = [WriteDouble];
388 let SchedRW = [Write64Bit] in {
389 // These instructions only exist on SI and CI
390 let SubtargetPredicate = isSICI, Predicates = [isSICI] in {
391 def V_LSHL_B64 : VOP3Inst <"v_lshl_b64", VOP3_Profile<VOP_PAT_GEN<VOP_I64_I64_I32>>, shl>;
392 def V_LSHR_B64 : VOP3Inst <"v_lshr_b64", VOP3_Profile<VOP_PAT_GEN<VOP_I64_I64_I32>>, srl>;
393 def V_ASHR_I64 : VOP3Inst <"v_ashr_i64", VOP3_Profile<VOP_PAT_GEN<VOP_I64_I64_I32>>, sra>;
394 def V_MULLIT_F32 : VOP3Inst <"v_mullit_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
395 } // End SubtargetPredicate = isSICI, Predicates = [isSICI]
397 let SubtargetPredicate = isVI in {
398 def V_LSHLREV_B64 : VOP3Inst <"v_lshlrev_b64", VOP3_Profile<VOP_I64_I32_I64>>;
399 def V_LSHRREV_B64 : VOP3Inst <"v_lshrrev_b64", VOP3_Profile<VOP_I64_I32_I64>>;
400 def V_ASHRREV_I64 : VOP3Inst <"v_ashrrev_i64", VOP3_Profile<VOP_I64_I32_I64>>;
401 } // End SubtargetPredicate = isVI
402 } // End SchedRW = [Write64Bit]
404 let Predicates = [isVI] in {
406 (getDivergentFrag<shl>.ret i64:$x, i32:$y),
407 (V_LSHLREV_B64 $y, $x)
410 (getDivergentFrag<srl>.ret i64:$x, i32:$y),
411 (V_LSHRREV_B64 $y, $x)
414 (getDivergentFrag<sra>.ret i64:$x, i32:$y),
415 (V_ASHRREV_I64 $y, $x)
420 let SubtargetPredicate = isCIVI in {
422 let Constraints = "@earlyclobber $vdst", SchedRW = [WriteQuarterRate32] in {
423 def V_QSAD_PK_U16_U8 : VOP3Inst <"v_qsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64, VOP3_CLAMP>>;
424 def V_MQSAD_U32_U8 : VOP3Inst <"v_mqsad_u32_u8", VOP3_Profile<VOP_V4I32_I64_I32_V4I32, VOP3_CLAMP>>;
425 } // End Constraints = "@earlyclobber $vdst", SchedRW = [WriteQuarterRate32]
427 let isCommutable = 1 in {
428 let SchedRW = [WriteQuarterRate32, WriteSALU] in {
429 def V_MAD_U64_U32 : VOP3Inst <"v_mad_u64_u32", VOP3b_I64_I1_I32_I32_I64>;
430 def V_MAD_I64_I32 : VOP3Inst <"v_mad_i64_i32", VOP3b_I64_I1_I32_I32_I64>;
431 } // End SchedRW = [WriteDouble, WriteSALU]
432 } // End isCommutable = 1
434 } // End SubtargetPredicate = isCIVI
437 def V_DIV_FIXUP_F16 : VOP3Inst <"v_div_fixup_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, AMDGPUdiv_fixup> {
438 let Predicates = [Has16BitInsts, isVIOnly];
439 let FPDPRounding = 1;
441 def V_DIV_FIXUP_F16_gfx9 : VOP3Inst <"v_div_fixup_f16_gfx9",
442 VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUdiv_fixup> {
443 let renamedInGFX9 = 1;
444 let Predicates = [Has16BitInsts, isGFX9];
445 let FPDPRounding = 1;
448 def V_FMA_F16 : VOP3Inst <"v_fma_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, fma> {
449 let Predicates = [Has16BitInsts, isVIOnly];
450 let FPDPRounding = 1;
452 def V_FMA_F16_gfx9 : VOP3Inst <"v_fma_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, fma> {
453 let renamedInGFX9 = 1;
454 let Predicates = [Has16BitInsts, isGFX9];
455 let FPDPRounding = 1;
458 let SubtargetPredicate = Has16BitInsts, isCommutable = 1 in {
460 let renamedInGFX9 = 1 in {
461 def V_MAD_U16 : VOP3Inst <"v_mad_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_CLAMP>>;
462 def V_MAD_I16 : VOP3Inst <"v_mad_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_CLAMP>>;
463 let FPDPRounding = 1 in {
464 def V_MAD_F16 : VOP3Inst <"v_mad_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, fmad>;
465 let Uses = [M0, EXEC] in {
466 def V_INTERP_P2_F16 : VOP3Interp <"v_interp_p2_f16", VOP3_INTERP16<[f16, f32, i32, f32]>>;
467 } // End Uses = [M0, EXEC]
468 } // End FPDPRounding = 1
469 } // End renamedInGFX9 = 1
471 let SubtargetPredicate = isGFX9 in {
472 def V_MAD_F16_gfx9 : VOP3Inst <"v_mad_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>> {
473 let FPDPRounding = 1;
475 def V_MAD_U16_gfx9 : VOP3Inst <"v_mad_u16_gfx9", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>>;
476 def V_MAD_I16_gfx9 : VOP3Inst <"v_mad_i16_gfx9", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>>;
477 def V_INTERP_P2_F16_gfx9 : VOP3Interp <"v_interp_p2_f16_gfx9", VOP3_INTERP16<[f16, f32, i32, f32]>>;
478 } // End SubtargetPredicate = isGFX9
480 let Uses = [M0, EXEC], FPDPRounding = 1 in {
481 def V_INTERP_P1LL_F16 : VOP3Interp <"v_interp_p1ll_f16", VOP3_INTERP16<[f32, f32, i32, untyped]>>;
482 def V_INTERP_P1LV_F16 : VOP3Interp <"v_interp_p1lv_f16", VOP3_INTERP16<[f32, f32, i32, f16]>>;
483 } // End Uses = [M0, EXEC], FPDPRounding = 1
485 } // End SubtargetPredicate = Has16BitInsts, isCommutable = 1
487 let SubtargetPredicate = isVI in {
488 def V_INTERP_P1_F32_e64 : VOP3Interp <"v_interp_p1_f32", VOP3_INTERP>;
489 def V_INTERP_P2_F32_e64 : VOP3Interp <"v_interp_p2_f32", VOP3_INTERP>;
490 def V_INTERP_MOV_F32_e64 : VOP3Interp <"v_interp_mov_f32", VOP3_INTERP_MOV>;
492 def V_PERM_B32 : VOP3Inst <"v_perm_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUperm>;
493 } // End SubtargetPredicate = isVI
495 let Predicates = [Has16BitInsts] in {
497 multiclass Ternary_i16_Pats <SDPatternOperator op1, SDPatternOperator op2,
498 Instruction inst, SDPatternOperator op3> {
500 (op2 (op1 i16:$src0, i16:$src1), i16:$src2),
501 (inst i16:$src0, i16:$src1, i16:$src2, (i1 0))
506 defm: Ternary_i16_Pats<mul, add, V_MAD_U16, zext>;
507 defm: Ternary_i16_Pats<mul, add, V_MAD_I16, sext>;
509 } // End Predicates = [Has16BitInsts]
511 class ThreeOpFrag<SDPatternOperator op1, SDPatternOperator op2> : PatFrag<
512 (ops node:$x, node:$y, node:$z),
513 // When the inner operation is used multiple times, selecting 3-op
514 // instructions may still be beneficial -- if the other users can be
515 // combined similarly. Let's be conservative for now.
516 (op2 (HasOneUseBinOp<op1> node:$x, node:$y), node:$z),
518 // Only use VALU ops when the result is divergent.
519 if (!N->isDivergent())
522 // Check constant bus limitations.
524 // Note: Use !isDivergent as a conservative proxy for whether the value
525 // is in an SGPR (uniform values can end up in VGPRs as well).
526 unsigned ConstantBusUses = 0;
527 for (unsigned i = 0; i < 3; ++i) {
528 if (!Operands[i]->isDivergent() &&
529 !isInlineImmediate(Operands[i].getNode())) {
531 if (ConstantBusUses >= 2)
539 let PredicateCodeUsesOperands = 1;
542 let SubtargetPredicate = isGFX9 in {
543 def V_PACK_B32_F16 : VOP3Inst <"v_pack_b32_f16", VOP3_Profile<VOP_B32_F16_F16, VOP3_OPSEL>>;
544 def V_LSHL_ADD_U32 : VOP3Inst <"v_lshl_add_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
545 def V_ADD_LSHL_U32 : VOP3Inst <"v_add_lshl_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
546 def V_ADD3_U32 : VOP3Inst <"v_add3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
547 def V_LSHL_OR_B32 : VOP3Inst <"v_lshl_or_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
548 def V_AND_OR_B32 : VOP3Inst <"v_and_or_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
549 def V_OR3_B32 : VOP3Inst <"v_or3_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
551 def V_XAD_U32 : VOP3Inst <"v_xad_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
553 def V_MED3_F16 : VOP3Inst <"v_med3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmed3>;
554 def V_MED3_I16 : VOP3Inst <"v_med3_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUsmed3>;
555 def V_MED3_U16 : VOP3Inst <"v_med3_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUumed3>;
557 def V_MIN3_F16 : VOP3Inst <"v_min3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmin3>;
558 def V_MIN3_I16 : VOP3Inst <"v_min3_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUsmin3>;
559 def V_MIN3_U16 : VOP3Inst <"v_min3_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUumin3>;
561 def V_MAX3_F16 : VOP3Inst <"v_max3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmax3>;
562 def V_MAX3_I16 : VOP3Inst <"v_max3_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUsmax3>;
563 def V_MAX3_U16 : VOP3Inst <"v_max3_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUumax3>;
565 def V_ADD_I16 : VOP3Inst <"v_add_i16", VOP3_Profile<VOP_I16_I16_I16, VOP3_OPSEL>>;
566 def V_SUB_I16 : VOP3Inst <"v_sub_i16", VOP3_Profile<VOP_I16_I16_I16, VOP3_OPSEL>>;
568 def V_MAD_U32_U16 : VOP3Inst <"v_mad_u32_u16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>;
569 def V_MAD_I32_I16 : VOP3Inst <"v_mad_i32_i16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>;
571 def V_CVT_PKNORM_I16_F16 : VOP3Inst <"v_cvt_pknorm_i16_f16", VOP3_Profile<VOP_B32_F16_F16, VOP3_OPSEL>>;
572 def V_CVT_PKNORM_U16_F16 : VOP3Inst <"v_cvt_pknorm_u16_f16", VOP3_Profile<VOP_B32_F16_F16, VOP3_OPSEL>>;
574 def V_ADD_I32_gfx9 : VOP3Inst <"v_add_i32_gfx9", VOP3_Profile<VOP_I32_I32_I32>>;
575 def V_SUB_I32_gfx9 : VOP3Inst <"v_sub_i32_gfx9", VOP3_Profile<VOP_I32_I32_I32>>;
578 class ThreeOp_i32_Pats <SDPatternOperator op1, SDPatternOperator op2, Instruction inst> : GCNPat <
579 // This matches (op2 (op1 i32:$src0, i32:$src1), i32:$src2) with conditions.
580 (ThreeOpFrag<op1, op2> i32:$src0, i32:$src1, i32:$src2),
581 (inst i32:$src0, i32:$src1, i32:$src2)
584 def : ThreeOp_i32_Pats<shl, add, V_LSHL_ADD_U32>;
585 def : ThreeOp_i32_Pats<add, shl, V_ADD_LSHL_U32>;
586 def : ThreeOp_i32_Pats<add, add, V_ADD3_U32>;
587 def : ThreeOp_i32_Pats<shl, or, V_LSHL_OR_B32>;
588 def : ThreeOp_i32_Pats<and, or, V_AND_OR_B32>;
589 def : ThreeOp_i32_Pats<or, or, V_OR3_B32>;
590 def : ThreeOp_i32_Pats<xor, add, V_XAD_U32>;
592 } // End SubtargetPredicate = isGFX9
594 //===----------------------------------------------------------------------===//
595 // Integer Clamp Patterns
596 //===----------------------------------------------------------------------===//
598 class getClampPat<VOPProfile P, SDPatternOperator node> {
599 dag ret3 = (P.DstVT (node P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2));
600 dag ret2 = (P.DstVT (node P.Src0VT:$src0, P.Src1VT:$src1));
601 dag ret1 = (P.DstVT (node P.Src0VT:$src0));
602 dag ret = !if(!eq(P.NumSrcArgs, 3), ret3,
603 !if(!eq(P.NumSrcArgs, 2), ret2,
607 class getClampRes<VOPProfile P, Instruction inst> {
608 dag ret3 = (inst P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2, (i1 0));
609 dag ret2 = (inst P.Src0VT:$src0, P.Src1VT:$src1, (i1 0));
610 dag ret1 = (inst P.Src0VT:$src0, (i1 0));
611 dag ret = !if(!eq(P.NumSrcArgs, 3), ret3,
612 !if(!eq(P.NumSrcArgs, 2), ret2,
616 class IntClampPat<VOP3Inst inst, SDPatternOperator node> : GCNPat<
617 getClampPat<inst.Pfl, node>.ret,
618 getClampRes<inst.Pfl, inst>.ret
621 def : IntClampPat<V_MAD_I32_I24, AMDGPUmad_i24>;
622 def : IntClampPat<V_MAD_U32_U24, AMDGPUmad_u24>;
624 def : IntClampPat<V_SAD_U8, int_amdgcn_sad_u8>;
625 def : IntClampPat<V_SAD_HI_U8, int_amdgcn_sad_hi_u8>;
626 def : IntClampPat<V_SAD_U16, int_amdgcn_sad_u16>;
628 def : IntClampPat<V_MSAD_U8, int_amdgcn_msad_u8>;
629 def : IntClampPat<V_MQSAD_PK_U16_U8, int_amdgcn_mqsad_pk_u16_u8>;
631 def : IntClampPat<V_QSAD_PK_U16_U8, int_amdgcn_qsad_pk_u16_u8>;
632 def : IntClampPat<V_MQSAD_U32_U8, int_amdgcn_mqsad_u32_u8>;
634 //===----------------------------------------------------------------------===//
636 //===----------------------------------------------------------------------===//
638 //===----------------------------------------------------------------------===//
640 //===----------------------------------------------------------------------===//
642 let AssemblerPredicates = [isSICI], DecoderNamespace = "SICI" in {
644 multiclass VOP3_Real_si<bits<9> op> {
645 def _si : VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.SI>,
646 VOP3e_si <op, !cast<VOP3_Pseudo>(NAME).Pfl>;
649 multiclass VOP3be_Real_si<bits<9> op> {
650 def _si : VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.SI>,
651 VOP3be_si <op, !cast<VOP3_Pseudo>(NAME).Pfl>;
654 } // End AssemblerPredicates = [isSICI], DecoderNamespace = "SICI"
656 defm V_MAD_LEGACY_F32 : VOP3_Real_si <0x140>;
657 defm V_MAD_F32 : VOP3_Real_si <0x141>;
658 defm V_MAD_I32_I24 : VOP3_Real_si <0x142>;
659 defm V_MAD_U32_U24 : VOP3_Real_si <0x143>;
660 defm V_CUBEID_F32 : VOP3_Real_si <0x144>;
661 defm V_CUBESC_F32 : VOP3_Real_si <0x145>;
662 defm V_CUBETC_F32 : VOP3_Real_si <0x146>;
663 defm V_CUBEMA_F32 : VOP3_Real_si <0x147>;
664 defm V_BFE_U32 : VOP3_Real_si <0x148>;
665 defm V_BFE_I32 : VOP3_Real_si <0x149>;
666 defm V_BFI_B32 : VOP3_Real_si <0x14a>;
667 defm V_FMA_F32 : VOP3_Real_si <0x14b>;
668 defm V_FMA_F64 : VOP3_Real_si <0x14c>;
669 defm V_LERP_U8 : VOP3_Real_si <0x14d>;
670 defm V_ALIGNBIT_B32 : VOP3_Real_si <0x14e>;
671 defm V_ALIGNBYTE_B32 : VOP3_Real_si <0x14f>;
672 defm V_MULLIT_F32 : VOP3_Real_si <0x150>;
673 defm V_MIN3_F32 : VOP3_Real_si <0x151>;
674 defm V_MIN3_I32 : VOP3_Real_si <0x152>;
675 defm V_MIN3_U32 : VOP3_Real_si <0x153>;
676 defm V_MAX3_F32 : VOP3_Real_si <0x154>;
677 defm V_MAX3_I32 : VOP3_Real_si <0x155>;
678 defm V_MAX3_U32 : VOP3_Real_si <0x156>;
679 defm V_MED3_F32 : VOP3_Real_si <0x157>;
680 defm V_MED3_I32 : VOP3_Real_si <0x158>;
681 defm V_MED3_U32 : VOP3_Real_si <0x159>;
682 defm V_SAD_U8 : VOP3_Real_si <0x15a>;
683 defm V_SAD_HI_U8 : VOP3_Real_si <0x15b>;
684 defm V_SAD_U16 : VOP3_Real_si <0x15c>;
685 defm V_SAD_U32 : VOP3_Real_si <0x15d>;
686 defm V_CVT_PK_U8_F32 : VOP3_Real_si <0x15e>;
687 defm V_DIV_FIXUP_F32 : VOP3_Real_si <0x15f>;
688 defm V_DIV_FIXUP_F64 : VOP3_Real_si <0x160>;
689 defm V_LSHL_B64 : VOP3_Real_si <0x161>;
690 defm V_LSHR_B64 : VOP3_Real_si <0x162>;
691 defm V_ASHR_I64 : VOP3_Real_si <0x163>;
692 defm V_ADD_F64 : VOP3_Real_si <0x164>;
693 defm V_MUL_F64 : VOP3_Real_si <0x165>;
694 defm V_MIN_F64 : VOP3_Real_si <0x166>;
695 defm V_MAX_F64 : VOP3_Real_si <0x167>;
696 defm V_LDEXP_F64 : VOP3_Real_si <0x168>;
697 defm V_MUL_LO_U32 : VOP3_Real_si <0x169>;
698 defm V_MUL_HI_U32 : VOP3_Real_si <0x16a>;
699 defm V_MUL_LO_I32 : VOP3_Real_si <0x16b>;
700 defm V_MUL_HI_I32 : VOP3_Real_si <0x16c>;
701 defm V_DIV_SCALE_F32 : VOP3be_Real_si <0x16d>;
702 defm V_DIV_SCALE_F64 : VOP3be_Real_si <0x16e>;
703 defm V_DIV_FMAS_F32 : VOP3_Real_si <0x16f>;
704 defm V_DIV_FMAS_F64 : VOP3_Real_si <0x170>;
705 defm V_MSAD_U8 : VOP3_Real_si <0x171>;
706 defm V_MQSAD_PK_U16_U8 : VOP3_Real_si <0x173>;
707 defm V_TRIG_PREOP_F64 : VOP3_Real_si <0x174>;
709 //===----------------------------------------------------------------------===//
711 //===----------------------------------------------------------------------===//
713 multiclass VOP3_Real_ci<bits<9> op> {
714 def _ci : VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.SI>,
715 VOP3e_si <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
716 let AssemblerPredicates = [isCIOnly];
717 let DecoderNamespace = "CI";
721 multiclass VOP3be_Real_ci<bits<9> op> {
722 def _ci : VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.SI>,
723 VOP3be_si <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
724 let AssemblerPredicates = [isCIOnly];
725 let DecoderNamespace = "CI";
729 defm V_QSAD_PK_U16_U8 : VOP3_Real_ci <0x172>;
730 defm V_MQSAD_U32_U8 : VOP3_Real_ci <0x175>;
731 defm V_MAD_U64_U32 : VOP3be_Real_ci <0x176>;
732 defm V_MAD_I64_I32 : VOP3be_Real_ci <0x177>;
734 //===----------------------------------------------------------------------===//
736 //===----------------------------------------------------------------------===//
738 let AssemblerPredicates = [isVI], DecoderNamespace = "VI" in {
740 multiclass VOP3_Real_vi<bits<10> op> {
741 def _vi : VOP3_Real<!cast<VOP_Pseudo>(NAME), SIEncodingFamily.VI>,
742 VOP3e_vi <op, !cast<VOP_Pseudo>(NAME).Pfl>;
745 multiclass VOP3be_Real_vi<bits<10> op> {
746 def _vi : VOP3_Real<!cast<VOP_Pseudo>(NAME), SIEncodingFamily.VI>,
747 VOP3be_vi <op, !cast<VOP_Pseudo>(NAME).Pfl>;
750 multiclass VOP3OpSel_Real_gfx9<bits<10> op> {
751 def _vi : VOP3_Real<!cast<VOP_Pseudo>(NAME), SIEncodingFamily.VI>,
752 VOP3OpSel_gfx9 <op, !cast<VOP_Pseudo>(NAME).Pfl>;
755 multiclass VOP3Interp_Real_vi<bits<10> op> {
756 def _vi : VOP3_Real<!cast<VOP_Pseudo>(NAME), SIEncodingFamily.VI>,
757 VOP3Interp_vi <op, !cast<VOP_Pseudo>(NAME).Pfl>;
760 } // End AssemblerPredicates = [isVI], DecoderNamespace = "VI"
762 let AssemblerPredicates = [isVIOnly], DecoderNamespace = "VI" in {
764 multiclass VOP3_F16_Real_vi<bits<10> op> {
765 def _vi : VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.VI>,
766 VOP3e_vi <op, !cast<VOP3_Pseudo>(NAME).Pfl>;
769 multiclass VOP3Interp_F16_Real_vi<bits<10> op> {
770 def _vi : VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.VI>,
771 VOP3Interp_vi <op, !cast<VOP3_Pseudo>(NAME).Pfl>;
774 } // End AssemblerPredicates = [isVIOnly], DecoderNamespace = "VI"
776 let AssemblerPredicates = [isGFX9], DecoderNamespace = "GFX9" in {
778 multiclass VOP3_F16_Real_gfx9<bits<10> op, string OpName, string AsmName> {
779 def _gfx9 : VOP3_Real<!cast<VOP3_Pseudo>(OpName), SIEncodingFamily.GFX9>,
780 VOP3e_vi <op, !cast<VOP3_Pseudo>(OpName).Pfl> {
781 VOP3_Pseudo ps = !cast<VOP3_Pseudo>(OpName);
782 let AsmString = AsmName # ps.AsmOperands;
786 multiclass VOP3OpSel_F16_Real_gfx9<bits<10> op, string AsmName> {
787 def _gfx9 : VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.GFX9>,
788 VOP3OpSel_gfx9 <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
789 VOP3_Pseudo ps = !cast<VOP3_Pseudo>(NAME);
790 let AsmString = AsmName # ps.AsmOperands;
794 multiclass VOP3Interp_F16_Real_gfx9<bits<10> op, string OpName, string AsmName> {
795 def _gfx9 : VOP3_Real<!cast<VOP3_Pseudo>(OpName), SIEncodingFamily.GFX9>,
796 VOP3Interp_vi <op, !cast<VOP3_Pseudo>(OpName).Pfl> {
797 VOP3_Pseudo ps = !cast<VOP3_Pseudo>(OpName);
798 let AsmString = AsmName # ps.AsmOperands;
802 multiclass VOP3_Real_gfx9<bits<10> op, string AsmName> {
803 def _gfx9 : VOP3_Real<!cast<VOP_Pseudo>(NAME), SIEncodingFamily.GFX9>,
804 VOP3e_vi <op, !cast<VOP_Pseudo>(NAME).Pfl> {
805 VOP_Pseudo ps = !cast<VOP_Pseudo>(NAME);
806 let AsmString = AsmName # ps.AsmOperands;
810 } // End AssemblerPredicates = [isGFX9], DecoderNamespace = "GFX9"
812 defm V_MAD_U64_U32 : VOP3be_Real_vi <0x1E8>;
813 defm V_MAD_I64_I32 : VOP3be_Real_vi <0x1E9>;
815 defm V_MAD_LEGACY_F32 : VOP3_Real_vi <0x1c0>;
816 defm V_MAD_F32 : VOP3_Real_vi <0x1c1>;
817 defm V_MAD_I32_I24 : VOP3_Real_vi <0x1c2>;
818 defm V_MAD_U32_U24 : VOP3_Real_vi <0x1c3>;
819 defm V_CUBEID_F32 : VOP3_Real_vi <0x1c4>;
820 defm V_CUBESC_F32 : VOP3_Real_vi <0x1c5>;
821 defm V_CUBETC_F32 : VOP3_Real_vi <0x1c6>;
822 defm V_CUBEMA_F32 : VOP3_Real_vi <0x1c7>;
823 defm V_BFE_U32 : VOP3_Real_vi <0x1c8>;
824 defm V_BFE_I32 : VOP3_Real_vi <0x1c9>;
825 defm V_BFI_B32 : VOP3_Real_vi <0x1ca>;
826 defm V_FMA_F32 : VOP3_Real_vi <0x1cb>;
827 defm V_FMA_F64 : VOP3_Real_vi <0x1cc>;
828 defm V_LERP_U8 : VOP3_Real_vi <0x1cd>;
829 defm V_ALIGNBIT_B32 : VOP3_Real_vi <0x1ce>;
830 defm V_ALIGNBYTE_B32 : VOP3_Real_vi <0x1cf>;
831 defm V_MIN3_F32 : VOP3_Real_vi <0x1d0>;
832 defm V_MIN3_I32 : VOP3_Real_vi <0x1d1>;
833 defm V_MIN3_U32 : VOP3_Real_vi <0x1d2>;
834 defm V_MAX3_F32 : VOP3_Real_vi <0x1d3>;
835 defm V_MAX3_I32 : VOP3_Real_vi <0x1d4>;
836 defm V_MAX3_U32 : VOP3_Real_vi <0x1d5>;
837 defm V_MED3_F32 : VOP3_Real_vi <0x1d6>;
838 defm V_MED3_I32 : VOP3_Real_vi <0x1d7>;
839 defm V_MED3_U32 : VOP3_Real_vi <0x1d8>;
840 defm V_SAD_U8 : VOP3_Real_vi <0x1d9>;
841 defm V_SAD_HI_U8 : VOP3_Real_vi <0x1da>;
842 defm V_SAD_U16 : VOP3_Real_vi <0x1db>;
843 defm V_SAD_U32 : VOP3_Real_vi <0x1dc>;
844 defm V_CVT_PK_U8_F32 : VOP3_Real_vi <0x1dd>;
845 defm V_DIV_FIXUP_F32 : VOP3_Real_vi <0x1de>;
846 defm V_DIV_FIXUP_F64 : VOP3_Real_vi <0x1df>;
847 defm V_DIV_SCALE_F32 : VOP3be_Real_vi <0x1e0>;
848 defm V_DIV_SCALE_F64 : VOP3be_Real_vi <0x1e1>;
849 defm V_DIV_FMAS_F32 : VOP3_Real_vi <0x1e2>;
850 defm V_DIV_FMAS_F64 : VOP3_Real_vi <0x1e3>;
851 defm V_MSAD_U8 : VOP3_Real_vi <0x1e4>;
852 defm V_QSAD_PK_U16_U8 : VOP3_Real_vi <0x1e5>;
853 defm V_MQSAD_PK_U16_U8 : VOP3_Real_vi <0x1e6>;
854 defm V_MQSAD_U32_U8 : VOP3_Real_vi <0x1e7>;
856 defm V_PERM_B32 : VOP3_Real_vi <0x1ed>;
858 defm V_MAD_F16 : VOP3_F16_Real_vi <0x1ea>;
859 defm V_MAD_U16 : VOP3_F16_Real_vi <0x1eb>;
860 defm V_MAD_I16 : VOP3_F16_Real_vi <0x1ec>;
861 defm V_FMA_F16 : VOP3_F16_Real_vi <0x1ee>;
862 defm V_DIV_FIXUP_F16 : VOP3_F16_Real_vi <0x1ef>;
863 defm V_INTERP_P2_F16 : VOP3Interp_F16_Real_vi <0x276>;
865 let FPDPRounding = 1 in {
866 defm V_MAD_LEGACY_F16 : VOP3_F16_Real_gfx9 <0x1ea, "V_MAD_F16", "v_mad_legacy_f16">;
867 defm V_FMA_LEGACY_F16 : VOP3_F16_Real_gfx9 <0x1ee, "V_FMA_F16", "v_fma_legacy_f16">;
868 defm V_DIV_FIXUP_LEGACY_F16 : VOP3_F16_Real_gfx9 <0x1ef, "V_DIV_FIXUP_F16", "v_div_fixup_legacy_f16">;
869 defm V_INTERP_P2_LEGACY_F16 : VOP3Interp_F16_Real_gfx9 <0x276, "V_INTERP_P2_F16", "v_interp_p2_legacy_f16">;
870 } // End FPDPRounding = 1
872 defm V_MAD_LEGACY_U16 : VOP3_F16_Real_gfx9 <0x1eb, "V_MAD_U16", "v_mad_legacy_u16">;
873 defm V_MAD_LEGACY_I16 : VOP3_F16_Real_gfx9 <0x1ec, "V_MAD_I16", "v_mad_legacy_i16">;
875 defm V_MAD_F16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x203, "v_mad_f16">;
876 defm V_MAD_U16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x204, "v_mad_u16">;
877 defm V_MAD_I16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x205, "v_mad_i16">;
878 defm V_FMA_F16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x206, "v_fma_f16">;
879 defm V_DIV_FIXUP_F16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x207, "v_div_fixup_f16">;
880 defm V_INTERP_P2_F16_gfx9 : VOP3Interp_F16_Real_gfx9 <0x277, "V_INTERP_P2_F16_gfx9", "v_interp_p2_f16">;
882 defm V_ADD_I32_gfx9 : VOP3_Real_gfx9 <0x29c, "v_add_i32">;
883 defm V_SUB_I32_gfx9 : VOP3_Real_gfx9 <0x29d, "v_sub_i32">;
885 defm V_INTERP_P1_F32_e64 : VOP3Interp_Real_vi <0x270>;
886 defm V_INTERP_P2_F32_e64 : VOP3Interp_Real_vi <0x271>;
887 defm V_INTERP_MOV_F32_e64 : VOP3Interp_Real_vi <0x272>;
889 defm V_INTERP_P1LL_F16 : VOP3Interp_Real_vi <0x274>;
890 defm V_INTERP_P1LV_F16 : VOP3Interp_Real_vi <0x275>;
891 defm V_ADD_F64 : VOP3_Real_vi <0x280>;
892 defm V_MUL_F64 : VOP3_Real_vi <0x281>;
893 defm V_MIN_F64 : VOP3_Real_vi <0x282>;
894 defm V_MAX_F64 : VOP3_Real_vi <0x283>;
895 defm V_LDEXP_F64 : VOP3_Real_vi <0x284>;
896 defm V_MUL_LO_U32 : VOP3_Real_vi <0x285>;
898 // removed from VI as identical to V_MUL_LO_U32
899 let isAsmParserOnly = 1 in {
900 defm V_MUL_LO_I32 : VOP3_Real_vi <0x285>;
903 defm V_MUL_HI_U32 : VOP3_Real_vi <0x286>;
904 defm V_MUL_HI_I32 : VOP3_Real_vi <0x287>;
906 defm V_READLANE_B32 : VOP3_Real_vi <0x289>;
907 defm V_WRITELANE_B32 : VOP3_Real_vi <0x28a>;
909 defm V_LSHLREV_B64 : VOP3_Real_vi <0x28f>;
910 defm V_LSHRREV_B64 : VOP3_Real_vi <0x290>;
911 defm V_ASHRREV_I64 : VOP3_Real_vi <0x291>;
912 defm V_TRIG_PREOP_F64 : VOP3_Real_vi <0x292>;
914 defm V_LSHL_ADD_U32 : VOP3_Real_vi <0x1fd>;
915 defm V_ADD_LSHL_U32 : VOP3_Real_vi <0x1fe>;
916 defm V_ADD3_U32 : VOP3_Real_vi <0x1ff>;
917 defm V_LSHL_OR_B32 : VOP3_Real_vi <0x200>;
918 defm V_AND_OR_B32 : VOP3_Real_vi <0x201>;
919 defm V_OR3_B32 : VOP3_Real_vi <0x202>;
920 defm V_PACK_B32_F16 : VOP3OpSel_Real_gfx9 <0x2a0>;
922 defm V_XAD_U32 : VOP3_Real_vi <0x1f3>;
924 defm V_MIN3_F16 : VOP3OpSel_Real_gfx9 <0x1f4>;
925 defm V_MIN3_I16 : VOP3OpSel_Real_gfx9 <0x1f5>;
926 defm V_MIN3_U16 : VOP3OpSel_Real_gfx9 <0x1f6>;
928 defm V_MAX3_F16 : VOP3OpSel_Real_gfx9 <0x1f7>;
929 defm V_MAX3_I16 : VOP3OpSel_Real_gfx9 <0x1f8>;
930 defm V_MAX3_U16 : VOP3OpSel_Real_gfx9 <0x1f9>;
932 defm V_MED3_F16 : VOP3OpSel_Real_gfx9 <0x1fa>;
933 defm V_MED3_I16 : VOP3OpSel_Real_gfx9 <0x1fb>;
934 defm V_MED3_U16 : VOP3OpSel_Real_gfx9 <0x1fc>;
936 defm V_ADD_I16 : VOP3OpSel_Real_gfx9 <0x29e>;
937 defm V_SUB_I16 : VOP3OpSel_Real_gfx9 <0x29f>;
939 defm V_MAD_U32_U16 : VOP3OpSel_Real_gfx9 <0x1f1>;
940 defm V_MAD_I32_I16 : VOP3OpSel_Real_gfx9 <0x1f2>;
942 defm V_CVT_PKNORM_I16_F16 : VOP3OpSel_Real_gfx9 <0x299>;
943 defm V_CVT_PKNORM_U16_F16 : VOP3OpSel_Real_gfx9 <0x29a>;