296 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck --check-prefix=GCN %s2 3; GCN-LABEL: {{^}}load_fma_store4; GCN-NOT: scratch_store5; ScratchSize: 0{{$}}6 7define amdgpu_kernel void @load_fma_store(ptr addrspace(3) nocapture readonly %arg, ptr addrspace(1) nocapture %arg1, i1 %cc) {8bb:9 %vgpr0 = call <32 x i32> asm sideeffect "; def $0","=${v[0:31]}"()10 %vgpr1 = call <32 x i32> asm sideeffect "; def $0","=${v[32:63]}"()11 %vgpr2 = call <32 x i32> asm sideeffect "; def $0","=${v[64:95]}"()12 %vgpr3 = call <32 x i32> asm sideeffect "; def $0","=${v[96:127]}"()13 br i1 %cc, label %bb1, label %bb214 15bb1:16 %tmp = getelementptr inbounds float, ptr addrspace(3) %arg, i32 117 %tmp2 = load float, ptr addrspace(3) %tmp, align 418 %tmp3 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 219 %tmp4 = load float, ptr addrspace(3) %tmp3, align 420 %tmp5 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 321 %tmp6 = load float, ptr addrspace(3) %tmp5, align 422 %tmp7 = tail call float @llvm.fmuladd.f32(float %tmp2, float %tmp4, float %tmp6)23 %tmp8 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 524 %tmp9 = load float, ptr addrspace(3) %tmp8, align 425 %tmp10 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 626 %tmp11 = load float, ptr addrspace(3) %tmp10, align 427 %tmp12 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 728 %tmp13 = load float, ptr addrspace(3) %tmp12, align 429 %tmp14 = tail call float @llvm.fmuladd.f32(float %tmp9, float %tmp11, float %tmp13)30 %tmp15 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 931 %tmp16 = load float, ptr addrspace(3) %tmp15, align 432 %tmp17 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1033 %tmp18 = load float, ptr addrspace(3) %tmp17, align 434 %tmp19 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1135 %tmp20 = load float, ptr addrspace(3) %tmp19, align 436 %tmp21 = tail call float @llvm.fmuladd.f32(float %tmp16, float %tmp18, float %tmp20)37 %tmp22 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1338 %tmp23 = load float, ptr addrspace(3) %tmp22, align 439 %tmp24 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1440 %tmp25 = load float, ptr addrspace(3) %tmp24, align 441 %tmp26 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1542 %tmp27 = load float, ptr addrspace(3) %tmp26, align 443 %tmp28 = tail call float @llvm.fmuladd.f32(float %tmp23, float %tmp25, float %tmp27)44 %tmp29 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1745 %tmp30 = load float, ptr addrspace(3) %tmp29, align 446 %tmp31 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1847 %tmp32 = load float, ptr addrspace(3) %tmp31, align 448 %tmp33 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1949 %tmp34 = load float, ptr addrspace(3) %tmp33, align 450 %tmp35 = tail call float @llvm.fmuladd.f32(float %tmp30, float %tmp32, float %tmp34)51 %tmp36 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2152 %tmp37 = load float, ptr addrspace(3) %tmp36, align 453 %tmp38 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2254 %tmp39 = load float, ptr addrspace(3) %tmp38, align 455 %tmp40 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2356 %tmp41 = load float, ptr addrspace(3) %tmp40, align 457 %tmp42 = tail call float @llvm.fmuladd.f32(float %tmp37, float %tmp39, float %tmp41)58 %tmp43 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2559 %tmp44 = load float, ptr addrspace(3) %tmp43, align 460 %tmp45 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2661 %tmp46 = load float, ptr addrspace(3) %tmp45, align 462 %tmp47 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2763 %tmp48 = load float, ptr addrspace(3) %tmp47, align 464 %tmp49 = tail call float @llvm.fmuladd.f32(float %tmp44, float %tmp46, float %tmp48)65 %tmp50 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2966 %tmp51 = load float, ptr addrspace(3) %tmp50, align 467 %tmp52 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3068 %tmp53 = load float, ptr addrspace(3) %tmp52, align 469 %tmp54 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3170 %tmp55 = load float, ptr addrspace(3) %tmp54, align 471 %tmp56 = tail call float @llvm.fmuladd.f32(float %tmp51, float %tmp53, float %tmp55)72 %tmp57 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3373 %tmp58 = load float, ptr addrspace(3) %tmp57, align 474 %tmp59 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3475 %tmp60 = load float, ptr addrspace(3) %tmp59, align 476 %tmp61 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3577 %tmp62 = load float, ptr addrspace(3) %tmp61, align 478 %tmp63 = tail call float @llvm.fmuladd.f32(float %tmp58, float %tmp60, float %tmp62)79 %tmp64 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3780 %tmp65 = load float, ptr addrspace(3) %tmp64, align 481 %tmp66 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3882 %tmp67 = load float, ptr addrspace(3) %tmp66, align 483 %tmp68 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3984 %tmp69 = load float, ptr addrspace(3) %tmp68, align 485 %tmp70 = tail call float @llvm.fmuladd.f32(float %tmp65, float %tmp67, float %tmp69)86 %tmp71 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4187 %tmp72 = load float, ptr addrspace(3) %tmp71, align 488 %tmp73 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4289 %tmp74 = load float, ptr addrspace(3) %tmp73, align 490 %tmp75 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4391 %tmp76 = load float, ptr addrspace(3) %tmp75, align 492 %tmp77 = tail call float @llvm.fmuladd.f32(float %tmp72, float %tmp74, float %tmp76)93 %tmp78 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4594 %tmp79 = load float, ptr addrspace(3) %tmp78, align 495 %tmp80 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4696 %tmp81 = load float, ptr addrspace(3) %tmp80, align 497 %tmp82 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4798 %tmp83 = load float, ptr addrspace(3) %tmp82, align 499 %tmp84 = tail call float @llvm.fmuladd.f32(float %tmp79, float %tmp81, float %tmp83)100 %tmp85 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 49101 %tmp86 = load float, ptr addrspace(3) %tmp85, align 4102 %tmp87 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 50103 %tmp88 = load float, ptr addrspace(3) %tmp87, align 4104 %tmp89 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 51105 %tmp90 = load float, ptr addrspace(3) %tmp89, align 4106 %tmp91 = tail call float @llvm.fmuladd.f32(float %tmp86, float %tmp88, float %tmp90)107 %tmp92 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 53108 %tmp93 = load float, ptr addrspace(3) %tmp92, align 4109 %tmp94 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 54110 %tmp95 = load float, ptr addrspace(3) %tmp94, align 4111 %tmp96 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 55112 %tmp97 = load float, ptr addrspace(3) %tmp96, align 4113 %tmp98 = tail call float @llvm.fmuladd.f32(float %tmp93, float %tmp95, float %tmp97)114 %tmp99 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 57115 %tmp100 = load float, ptr addrspace(3) %tmp99, align 4116 %tmp101 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 58117 %tmp102 = load float, ptr addrspace(3) %tmp101, align 4118 %tmp103 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 59119 %tmp104 = load float, ptr addrspace(3) %tmp103, align 4120 %tmp105 = tail call float @llvm.fmuladd.f32(float %tmp100, float %tmp102, float %tmp104)121 %tmp106 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 61122 %tmp107 = load float, ptr addrspace(3) %tmp106, align 4123 %tmp108 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 62124 %tmp109 = load float, ptr addrspace(3) %tmp108, align 4125 %tmp110 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 63126 %tmp111 = load float, ptr addrspace(3) %tmp110, align 4127 %tmp112 = tail call float @llvm.fmuladd.f32(float %tmp107, float %tmp109, float %tmp111)128 %tmp113 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 65129 %tmp114 = load float, ptr addrspace(3) %tmp113, align 4130 %tmp115 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 66131 %tmp116 = load float, ptr addrspace(3) %tmp115, align 4132 %tmp117 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 67133 %tmp118 = load float, ptr addrspace(3) %tmp117, align 4134 %tmp119 = tail call float @llvm.fmuladd.f32(float %tmp114, float %tmp116, float %tmp118)135 %tmp120 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 69136 %tmp121 = load float, ptr addrspace(3) %tmp120, align 4137 %tmp122 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 70138 %tmp123 = load float, ptr addrspace(3) %tmp122, align 4139 %tmp124 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 71140 %tmp125 = load float, ptr addrspace(3) %tmp124, align 4141 %tmp126 = tail call float @llvm.fmuladd.f32(float %tmp121, float %tmp123, float %tmp125)142 %tmp127 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 73143 %tmp128 = load float, ptr addrspace(3) %tmp127, align 4144 %tmp129 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 74145 %tmp130 = load float, ptr addrspace(3) %tmp129, align 4146 %tmp131 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 75147 %tmp132 = load float, ptr addrspace(3) %tmp131, align 4148 %tmp133 = tail call float @llvm.fmuladd.f32(float %tmp128, float %tmp130, float %tmp132)149 %tmp134 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 77150 %tmp135 = load float, ptr addrspace(3) %tmp134, align 4151 %tmp136 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 78152 %tmp137 = load float, ptr addrspace(3) %tmp136, align 4153 %tmp138 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 79154 %tmp139 = load float, ptr addrspace(3) %tmp138, align 4155 %tmp140 = tail call float @llvm.fmuladd.f32(float %tmp135, float %tmp137, float %tmp139)156 %tmp141 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 81157 %tmp142 = load float, ptr addrspace(3) %tmp141, align 4158 %tmp143 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 82159 %tmp144 = load float, ptr addrspace(3) %tmp143, align 4160 %tmp145 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 83161 %tmp146 = load float, ptr addrspace(3) %tmp145, align 4162 %tmp147 = tail call float @llvm.fmuladd.f32(float %tmp142, float %tmp144, float %tmp146)163 %tmp148 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 85164 %tmp149 = load float, ptr addrspace(3) %tmp148, align 4165 %tmp150 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 86166 %tmp151 = load float, ptr addrspace(3) %tmp150, align 4167 %tmp152 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 87168 %tmp153 = load float, ptr addrspace(3) %tmp152, align 4169 %tmp154 = tail call float @llvm.fmuladd.f32(float %tmp149, float %tmp151, float %tmp153)170 %tmp155 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 89171 %tmp156 = load float, ptr addrspace(3) %tmp155, align 4172 %tmp157 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 90173 %tmp158 = load float, ptr addrspace(3) %tmp157, align 4174 %tmp159 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 91175 %tmp160 = load float, ptr addrspace(3) %tmp159, align 4176 %tmp161 = tail call float @llvm.fmuladd.f32(float %tmp156, float %tmp158, float %tmp160)177 %tmp162 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 93178 %tmp163 = load float, ptr addrspace(3) %tmp162, align 4179 %tmp164 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 94180 %tmp165 = load float, ptr addrspace(3) %tmp164, align 4181 %tmp166 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 95182 %tmp167 = load float, ptr addrspace(3) %tmp166, align 4183 %tmp168 = tail call float @llvm.fmuladd.f32(float %tmp163, float %tmp165, float %tmp167)184 %tmp169 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 97185 %tmp170 = load float, ptr addrspace(3) %tmp169, align 4186 %tmp171 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 98187 %tmp172 = load float, ptr addrspace(3) %tmp171, align 4188 %tmp173 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 99189 %tmp174 = load float, ptr addrspace(3) %tmp173, align 4190 %tmp175 = tail call float @llvm.fmuladd.f32(float %tmp170, float %tmp172, float %tmp174)191 %tmp176 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 101192 %tmp177 = load float, ptr addrspace(3) %tmp176, align 4193 %tmp178 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 102194 %tmp179 = load float, ptr addrspace(3) %tmp178, align 4195 %tmp180 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 103196 %tmp181 = load float, ptr addrspace(3) %tmp180, align 4197 %tmp182 = tail call float @llvm.fmuladd.f32(float %tmp177, float %tmp179, float %tmp181)198 %tmp183 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 105199 %tmp184 = load float, ptr addrspace(3) %tmp183, align 4200 %tmp185 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 106201 %tmp186 = load float, ptr addrspace(3) %tmp185, align 4202 %tmp187 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 107203 %tmp188 = load float, ptr addrspace(3) %tmp187, align 4204 %tmp189 = tail call float @llvm.fmuladd.f32(float %tmp184, float %tmp186, float %tmp188)205 %tmp190 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 109206 %tmp191 = load float, ptr addrspace(3) %tmp190, align 4207 %tmp192 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 110208 %tmp193 = load float, ptr addrspace(3) %tmp192, align 4209 %tmp194 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 111210 %tmp195 = load float, ptr addrspace(3) %tmp194, align 4211 %tmp196 = tail call float @llvm.fmuladd.f32(float %tmp191, float %tmp193, float %tmp195)212 %tmp197 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 113213 %tmp198 = load float, ptr addrspace(3) %tmp197, align 4214 %tmp199 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 114215 %tmp200 = load float, ptr addrspace(3) %tmp199, align 4216 %tmp201 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 115217 %tmp202 = load float, ptr addrspace(3) %tmp201, align 4218 %tmp203 = tail call float @llvm.fmuladd.f32(float %tmp198, float %tmp200, float %tmp202)219 %tmp204 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 117220 %tmp205 = load float, ptr addrspace(3) %tmp204, align 4221 %tmp206 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 118222 %tmp207 = load float, ptr addrspace(3) %tmp206, align 4223 %tmp208 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 119224 %tmp209 = load float, ptr addrspace(3) %tmp208, align 4225 %tmp210 = tail call float @llvm.fmuladd.f32(float %tmp205, float %tmp207, float %tmp209)226 store float %tmp7, ptr addrspace(1) %arg1, align 4227 %tmp449 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 1228 store float %tmp14, ptr addrspace(1) %tmp449, align 4229 %tmp450 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 2230 store float %tmp21, ptr addrspace(1) %tmp450, align 4231 %tmp451 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 3232 store float %tmp28, ptr addrspace(1) %tmp451, align 4233 %tmp452 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 4234 store float %tmp35, ptr addrspace(1) %tmp452, align 4235 %tmp453 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 5236 store float %tmp42, ptr addrspace(1) %tmp453, align 4237 %tmp454 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 6238 store float %tmp49, ptr addrspace(1) %tmp454, align 4239 %tmp455 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 7240 store float %tmp56, ptr addrspace(1) %tmp455, align 4241 %tmp456 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 8242 store float %tmp63, ptr addrspace(1) %tmp456, align 4243 %tmp457 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 9244 store float %tmp70, ptr addrspace(1) %tmp457, align 4245 %tmp458 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 10246 store float %tmp77, ptr addrspace(1) %tmp458, align 4247 %tmp459 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 11248 store float %tmp84, ptr addrspace(1) %tmp459, align 4249 %tmp460 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 12250 store float %tmp91, ptr addrspace(1) %tmp460, align 4251 %tmp461 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 13252 store float %tmp98, ptr addrspace(1) %tmp461, align 4253 %tmp462 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 14254 store float %tmp105, ptr addrspace(1) %tmp462, align 4255 %tmp463 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 15256 store float %tmp112, ptr addrspace(1) %tmp463, align 4257 %tmp464 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 16258 store float %tmp119, ptr addrspace(1) %tmp464, align 4259 %tmp465 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 17260 store float %tmp126, ptr addrspace(1) %tmp465, align 4261 %tmp466 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 18262 store float %tmp133, ptr addrspace(1) %tmp466, align 4263 %tmp467 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 19264 store float %tmp140, ptr addrspace(1) %tmp467, align 4265 %tmp468 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 20266 store float %tmp147, ptr addrspace(1) %tmp468, align 4267 %tmp469 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 21268 store float %tmp154, ptr addrspace(1) %tmp469, align 4269 %tmp470 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 22270 store float %tmp161, ptr addrspace(1) %tmp470, align 4271 %tmp471 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 23272 store float %tmp168, ptr addrspace(1) %tmp471, align 4273 %tmp472 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 24274 store float %tmp175, ptr addrspace(1) %tmp472, align 4275 %tmp473 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 25276 store float %tmp182, ptr addrspace(1) %tmp473, align 4277 %tmp474 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 26278 store float %tmp189, ptr addrspace(1) %tmp474, align 4279 %tmp475 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 27280 store float %tmp196, ptr addrspace(1) %tmp475, align 4281 %tmp476 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 28282 store float %tmp203, ptr addrspace(1) %tmp476, align 4283 %tmp477 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 29284 store float %tmp210, ptr addrspace(1) %tmp477, align 4285 br i1 %cc, label %bb1, label %bb2286 287bb2:288 call void asm sideeffect "; use $0","{v[0:31]}"(<32 x i32> %vgpr0)289 call void asm sideeffect "; use $0","{v[0:31]}"(<32 x i32> %vgpr1)290 call void asm sideeffect "; use $0","{v[0:31]}"(<32 x i32> %vgpr2)291 call void asm sideeffect "; use $0","{v[0:31]}"(<32 x i32> %vgpr3)292 ret void293}294 295declare float @llvm.fmuladd.f32(float, float, float)296