602 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck --check-prefix=MISCHED %s2; RUN: llc -mtriple=amdgcn -mcpu=tonga -misched=gcn-iterative-ilp < %s | FileCheck --check-prefix=GCN-ILP %s3; RUN: llc -mtriple=amdgcn -mcpu=tonga -amdgpu-sched-strategy=iterative-ilp < %s | FileCheck --check-prefix=GCN-ILP %s4 5; Test the scheduler when only one wave is requested. The result should be high register usage and max ILP.6 7; We expect a three digit VGPR usage here since only one wave requested.8;9; GCN-ILP: NumVgprs: {{[0-9][0-9][0-9]$}}10 11; FIXME: The machine scheduler is doing a poor job at maximizing ILP here.12; However, if we had not requested only one wave register usage would indeed be13; much lower, demonstrating that is the purpose of this test.14;15; MISCHED: NumVgprs: {{[7-9][0-9]$}}16 17define amdgpu_kernel void @load_fma_store(ptr addrspace(3) nocapture readonly %arg, ptr addrspace(1) nocapture %arg1) #1 {18bb:19 %tmp = getelementptr inbounds float, ptr addrspace(3) %arg, i32 120 %tmp2 = load float, ptr addrspace(3) %tmp, align 421 %tmp3 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 222 %tmp4 = load float, ptr addrspace(3) %tmp3, align 423 %tmp5 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 324 %tmp6 = load float, ptr addrspace(3) %tmp5, align 425 %tmp7 = tail call float @llvm.fmuladd.f32(float %tmp2, float %tmp4, float %tmp6)26 %tmp8 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 527 %tmp9 = load float, ptr addrspace(3) %tmp8, align 428 %tmp10 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 629 %tmp11 = load float, ptr addrspace(3) %tmp10, align 430 %tmp12 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 731 %tmp13 = load float, ptr addrspace(3) %tmp12, align 432 %tmp14 = tail call float @llvm.fmuladd.f32(float %tmp9, float %tmp11, float %tmp13)33 %tmp15 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 934 %tmp16 = load float, ptr addrspace(3) %tmp15, align 435 %tmp17 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1036 %tmp18 = load float, ptr addrspace(3) %tmp17, align 437 %tmp19 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1138 %tmp20 = load float, ptr addrspace(3) %tmp19, align 439 %tmp21 = tail call float @llvm.fmuladd.f32(float %tmp16, float %tmp18, float %tmp20)40 %tmp22 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1341 %tmp23 = load float, ptr addrspace(3) %tmp22, align 442 %tmp24 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1443 %tmp25 = load float, ptr addrspace(3) %tmp24, align 444 %tmp26 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1545 %tmp27 = load float, ptr addrspace(3) %tmp26, align 446 %tmp28 = tail call float @llvm.fmuladd.f32(float %tmp23, float %tmp25, float %tmp27)47 %tmp29 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1748 %tmp30 = load float, ptr addrspace(3) %tmp29, align 449 %tmp31 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1850 %tmp32 = load float, ptr addrspace(3) %tmp31, align 451 %tmp33 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1952 %tmp34 = load float, ptr addrspace(3) %tmp33, align 453 %tmp35 = tail call float @llvm.fmuladd.f32(float %tmp30, float %tmp32, float %tmp34)54 %tmp36 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2155 %tmp37 = load float, ptr addrspace(3) %tmp36, align 456 %tmp38 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2257 %tmp39 = load float, ptr addrspace(3) %tmp38, align 458 %tmp40 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2359 %tmp41 = load float, ptr addrspace(3) %tmp40, align 460 %tmp42 = tail call float @llvm.fmuladd.f32(float %tmp37, float %tmp39, float %tmp41)61 %tmp43 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2562 %tmp44 = load float, ptr addrspace(3) %tmp43, align 463 %tmp45 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2664 %tmp46 = load float, ptr addrspace(3) %tmp45, align 465 %tmp47 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2766 %tmp48 = load float, ptr addrspace(3) %tmp47, align 467 %tmp49 = tail call float @llvm.fmuladd.f32(float %tmp44, float %tmp46, float %tmp48)68 %tmp50 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2969 %tmp51 = load float, ptr addrspace(3) %tmp50, align 470 %tmp52 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3071 %tmp53 = load float, ptr addrspace(3) %tmp52, align 472 %tmp54 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3173 %tmp55 = load float, ptr addrspace(3) %tmp54, align 474 %tmp56 = tail call float @llvm.fmuladd.f32(float %tmp51, float %tmp53, float %tmp55)75 %tmp57 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3376 %tmp58 = load float, ptr addrspace(3) %tmp57, align 477 %tmp59 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3478 %tmp60 = load float, ptr addrspace(3) %tmp59, align 479 %tmp61 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3580 %tmp62 = load float, ptr addrspace(3) %tmp61, align 481 %tmp63 = tail call float @llvm.fmuladd.f32(float %tmp58, float %tmp60, float %tmp62)82 %tmp64 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3783 %tmp65 = load float, ptr addrspace(3) %tmp64, align 484 %tmp66 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3885 %tmp67 = load float, ptr addrspace(3) %tmp66, align 486 %tmp68 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3987 %tmp69 = load float, ptr addrspace(3) %tmp68, align 488 %tmp70 = tail call float @llvm.fmuladd.f32(float %tmp65, float %tmp67, float %tmp69)89 %tmp71 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4190 %tmp72 = load float, ptr addrspace(3) %tmp71, align 491 %tmp73 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4292 %tmp74 = load float, ptr addrspace(3) %tmp73, align 493 %tmp75 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4394 %tmp76 = load float, ptr addrspace(3) %tmp75, align 495 %tmp77 = tail call float @llvm.fmuladd.f32(float %tmp72, float %tmp74, float %tmp76)96 %tmp78 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4597 %tmp79 = load float, ptr addrspace(3) %tmp78, align 498 %tmp80 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4699 %tmp81 = load float, ptr addrspace(3) %tmp80, align 4100 %tmp82 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 47101 %tmp83 = load float, ptr addrspace(3) %tmp82, align 4102 %tmp84 = tail call float @llvm.fmuladd.f32(float %tmp79, float %tmp81, float %tmp83)103 %tmp85 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 49104 %tmp86 = load float, ptr addrspace(3) %tmp85, align 4105 %tmp87 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 50106 %tmp88 = load float, ptr addrspace(3) %tmp87, align 4107 %tmp89 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 51108 %tmp90 = load float, ptr addrspace(3) %tmp89, align 4109 %tmp91 = tail call float @llvm.fmuladd.f32(float %tmp86, float %tmp88, float %tmp90)110 %tmp92 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 53111 %tmp93 = load float, ptr addrspace(3) %tmp92, align 4112 %tmp94 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 54113 %tmp95 = load float, ptr addrspace(3) %tmp94, align 4114 %tmp96 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 55115 %tmp97 = load float, ptr addrspace(3) %tmp96, align 4116 %tmp98 = tail call float @llvm.fmuladd.f32(float %tmp93, float %tmp95, float %tmp97)117 %tmp99 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 57118 %tmp100 = load float, ptr addrspace(3) %tmp99, align 4119 %tmp101 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 58120 %tmp102 = load float, ptr addrspace(3) %tmp101, align 4121 %tmp103 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 59122 %tmp104 = load float, ptr addrspace(3) %tmp103, align 4123 %tmp105 = tail call float @llvm.fmuladd.f32(float %tmp100, float %tmp102, float %tmp104)124 %tmp106 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 61125 %tmp107 = load float, ptr addrspace(3) %tmp106, align 4126 %tmp108 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 62127 %tmp109 = load float, ptr addrspace(3) %tmp108, align 4128 %tmp110 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 63129 %tmp111 = load float, ptr addrspace(3) %tmp110, align 4130 %tmp112 = tail call float @llvm.fmuladd.f32(float %tmp107, float %tmp109, float %tmp111)131 %tmp113 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 65132 %tmp114 = load float, ptr addrspace(3) %tmp113, align 4133 %tmp115 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 66134 %tmp116 = load float, ptr addrspace(3) %tmp115, align 4135 %tmp117 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 67136 %tmp118 = load float, ptr addrspace(3) %tmp117, align 4137 %tmp119 = tail call float @llvm.fmuladd.f32(float %tmp114, float %tmp116, float %tmp118)138 %tmp120 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 69139 %tmp121 = load float, ptr addrspace(3) %tmp120, align 4140 %tmp122 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 70141 %tmp123 = load float, ptr addrspace(3) %tmp122, align 4142 %tmp124 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 71143 %tmp125 = load float, ptr addrspace(3) %tmp124, align 4144 %tmp126 = tail call float @llvm.fmuladd.f32(float %tmp121, float %tmp123, float %tmp125)145 %tmp127 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 73146 %tmp128 = load float, ptr addrspace(3) %tmp127, align 4147 %tmp129 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 74148 %tmp130 = load float, ptr addrspace(3) %tmp129, align 4149 %tmp131 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 75150 %tmp132 = load float, ptr addrspace(3) %tmp131, align 4151 %tmp133 = tail call float @llvm.fmuladd.f32(float %tmp128, float %tmp130, float %tmp132)152 %tmp134 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 77153 %tmp135 = load float, ptr addrspace(3) %tmp134, align 4154 %tmp136 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 78155 %tmp137 = load float, ptr addrspace(3) %tmp136, align 4156 %tmp138 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 79157 %tmp139 = load float, ptr addrspace(3) %tmp138, align 4158 %tmp140 = tail call float @llvm.fmuladd.f32(float %tmp135, float %tmp137, float %tmp139)159 %tmp141 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 81160 %tmp142 = load float, ptr addrspace(3) %tmp141, align 4161 %tmp143 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 82162 %tmp144 = load float, ptr addrspace(3) %tmp143, align 4163 %tmp145 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 83164 %tmp146 = load float, ptr addrspace(3) %tmp145, align 4165 %tmp147 = tail call float @llvm.fmuladd.f32(float %tmp142, float %tmp144, float %tmp146)166 %tmp148 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 85167 %tmp149 = load float, ptr addrspace(3) %tmp148, align 4168 %tmp150 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 86169 %tmp151 = load float, ptr addrspace(3) %tmp150, align 4170 %tmp152 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 87171 %tmp153 = load float, ptr addrspace(3) %tmp152, align 4172 %tmp154 = tail call float @llvm.fmuladd.f32(float %tmp149, float %tmp151, float %tmp153)173 %tmp155 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 89174 %tmp156 = load float, ptr addrspace(3) %tmp155, align 4175 %tmp157 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 90176 %tmp158 = load float, ptr addrspace(3) %tmp157, align 4177 %tmp159 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 91178 %tmp160 = load float, ptr addrspace(3) %tmp159, align 4179 %tmp161 = tail call float @llvm.fmuladd.f32(float %tmp156, float %tmp158, float %tmp160)180 %tmp162 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 93181 %tmp163 = load float, ptr addrspace(3) %tmp162, align 4182 %tmp164 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 94183 %tmp165 = load float, ptr addrspace(3) %tmp164, align 4184 %tmp166 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 95185 %tmp167 = load float, ptr addrspace(3) %tmp166, align 4186 %tmp168 = tail call float @llvm.fmuladd.f32(float %tmp163, float %tmp165, float %tmp167)187 %tmp169 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 97188 %tmp170 = load float, ptr addrspace(3) %tmp169, align 4189 %tmp171 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 98190 %tmp172 = load float, ptr addrspace(3) %tmp171, align 4191 %tmp173 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 99192 %tmp174 = load float, ptr addrspace(3) %tmp173, align 4193 %tmp175 = tail call float @llvm.fmuladd.f32(float %tmp170, float %tmp172, float %tmp174)194 %tmp176 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 101195 %tmp177 = load float, ptr addrspace(3) %tmp176, align 4196 %tmp178 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 102197 %tmp179 = load float, ptr addrspace(3) %tmp178, align 4198 %tmp180 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 103199 %tmp181 = load float, ptr addrspace(3) %tmp180, align 4200 %tmp182 = tail call float @llvm.fmuladd.f32(float %tmp177, float %tmp179, float %tmp181)201 %tmp183 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 105202 %tmp184 = load float, ptr addrspace(3) %tmp183, align 4203 %tmp185 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 106204 %tmp186 = load float, ptr addrspace(3) %tmp185, align 4205 %tmp187 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 107206 %tmp188 = load float, ptr addrspace(3) %tmp187, align 4207 %tmp189 = tail call float @llvm.fmuladd.f32(float %tmp184, float %tmp186, float %tmp188)208 %tmp190 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 109209 %tmp191 = load float, ptr addrspace(3) %tmp190, align 4210 %tmp192 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 110211 %tmp193 = load float, ptr addrspace(3) %tmp192, align 4212 %tmp194 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 111213 %tmp195 = load float, ptr addrspace(3) %tmp194, align 4214 %tmp196 = tail call float @llvm.fmuladd.f32(float %tmp191, float %tmp193, float %tmp195)215 %tmp197 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 113216 %tmp198 = load float, ptr addrspace(3) %tmp197, align 4217 %tmp199 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 114218 %tmp200 = load float, ptr addrspace(3) %tmp199, align 4219 %tmp201 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 115220 %tmp202 = load float, ptr addrspace(3) %tmp201, align 4221 %tmp203 = tail call float @llvm.fmuladd.f32(float %tmp198, float %tmp200, float %tmp202)222 %tmp204 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 117223 %tmp205 = load float, ptr addrspace(3) %tmp204, align 4224 %tmp206 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 118225 %tmp207 = load float, ptr addrspace(3) %tmp206, align 4226 %tmp208 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 119227 %tmp209 = load float, ptr addrspace(3) %tmp208, align 4228 %tmp210 = tail call float @llvm.fmuladd.f32(float %tmp205, float %tmp207, float %tmp209)229 %tmp211 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 121230 %tmp212 = load float, ptr addrspace(3) %tmp211, align 4231 %tmp213 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 122232 %tmp214 = load float, ptr addrspace(3) %tmp213, align 4233 %tmp215 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 123234 %tmp216 = load float, ptr addrspace(3) %tmp215, align 4235 %tmp217 = tail call float @llvm.fmuladd.f32(float %tmp212, float %tmp214, float %tmp216)236 %tmp218 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 125237 %tmp219 = load float, ptr addrspace(3) %tmp218, align 4238 %tmp220 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 126239 %tmp221 = load float, ptr addrspace(3) %tmp220, align 4240 %tmp222 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 127241 %tmp223 = load float, ptr addrspace(3) %tmp222, align 4242 %tmp224 = tail call float @llvm.fmuladd.f32(float %tmp219, float %tmp221, float %tmp223)243 %tmp225 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 129244 %tmp226 = load float, ptr addrspace(3) %tmp225, align 4245 %tmp227 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 130246 %tmp228 = load float, ptr addrspace(3) %tmp227, align 4247 %tmp229 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 131248 %tmp230 = load float, ptr addrspace(3) %tmp229, align 4249 %tmp231 = tail call float @llvm.fmuladd.f32(float %tmp226, float %tmp228, float %tmp230)250 %tmp232 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 133251 %tmp233 = load float, ptr addrspace(3) %tmp232, align 4252 %tmp234 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 134253 %tmp235 = load float, ptr addrspace(3) %tmp234, align 4254 %tmp236 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 135255 %tmp237 = load float, ptr addrspace(3) %tmp236, align 4256 %tmp238 = tail call float @llvm.fmuladd.f32(float %tmp233, float %tmp235, float %tmp237)257 %tmp239 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 137258 %tmp240 = load float, ptr addrspace(3) %tmp239, align 4259 %tmp241 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 138260 %tmp242 = load float, ptr addrspace(3) %tmp241, align 4261 %tmp243 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 139262 %tmp244 = load float, ptr addrspace(3) %tmp243, align 4263 %tmp245 = tail call float @llvm.fmuladd.f32(float %tmp240, float %tmp242, float %tmp244)264 %tmp246 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 141265 %tmp247 = load float, ptr addrspace(3) %tmp246, align 4266 %tmp248 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 142267 %tmp249 = load float, ptr addrspace(3) %tmp248, align 4268 %tmp250 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 143269 %tmp251 = load float, ptr addrspace(3) %tmp250, align 4270 %tmp252 = tail call float @llvm.fmuladd.f32(float %tmp247, float %tmp249, float %tmp251)271 %tmp253 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 145272 %tmp254 = load float, ptr addrspace(3) %tmp253, align 4273 %tmp255 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 146274 %tmp256 = load float, ptr addrspace(3) %tmp255, align 4275 %tmp257 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 147276 %tmp258 = load float, ptr addrspace(3) %tmp257, align 4277 %tmp259 = tail call float @llvm.fmuladd.f32(float %tmp254, float %tmp256, float %tmp258)278 %tmp260 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 149279 %tmp261 = load float, ptr addrspace(3) %tmp260, align 4280 %tmp262 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 150281 %tmp263 = load float, ptr addrspace(3) %tmp262, align 4282 %tmp264 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 151283 %tmp265 = load float, ptr addrspace(3) %tmp264, align 4284 %tmp266 = tail call float @llvm.fmuladd.f32(float %tmp261, float %tmp263, float %tmp265)285 %tmp267 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 153286 %tmp268 = load float, ptr addrspace(3) %tmp267, align 4287 %tmp269 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 154288 %tmp270 = load float, ptr addrspace(3) %tmp269, align 4289 %tmp271 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 155290 %tmp272 = load float, ptr addrspace(3) %tmp271, align 4291 %tmp273 = tail call float @llvm.fmuladd.f32(float %tmp268, float %tmp270, float %tmp272)292 %tmp274 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 157293 %tmp275 = load float, ptr addrspace(3) %tmp274, align 4294 %tmp276 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 158295 %tmp277 = load float, ptr addrspace(3) %tmp276, align 4296 %tmp278 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 159297 %tmp279 = load float, ptr addrspace(3) %tmp278, align 4298 %tmp280 = tail call float @llvm.fmuladd.f32(float %tmp275, float %tmp277, float %tmp279)299 %tmp281 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 161300 %tmp282 = load float, ptr addrspace(3) %tmp281, align 4301 %tmp283 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 162302 %tmp284 = load float, ptr addrspace(3) %tmp283, align 4303 %tmp285 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 163304 %tmp286 = load float, ptr addrspace(3) %tmp285, align 4305 %tmp287 = tail call float @llvm.fmuladd.f32(float %tmp282, float %tmp284, float %tmp286)306 %tmp288 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 165307 %tmp289 = load float, ptr addrspace(3) %tmp288, align 4308 %tmp290 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 166309 %tmp291 = load float, ptr addrspace(3) %tmp290, align 4310 %tmp292 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 167311 %tmp293 = load float, ptr addrspace(3) %tmp292, align 4312 %tmp294 = tail call float @llvm.fmuladd.f32(float %tmp289, float %tmp291, float %tmp293)313 %tmp295 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 169314 %tmp296 = load float, ptr addrspace(3) %tmp295, align 4315 %tmp297 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 170316 %tmp298 = load float, ptr addrspace(3) %tmp297, align 4317 %tmp299 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 171318 %tmp300 = load float, ptr addrspace(3) %tmp299, align 4319 %tmp301 = tail call float @llvm.fmuladd.f32(float %tmp296, float %tmp298, float %tmp300)320 %tmp302 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 173321 %tmp303 = load float, ptr addrspace(3) %tmp302, align 4322 %tmp304 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 174323 %tmp305 = load float, ptr addrspace(3) %tmp304, align 4324 %tmp306 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 175325 %tmp307 = load float, ptr addrspace(3) %tmp306, align 4326 %tmp308 = tail call float @llvm.fmuladd.f32(float %tmp303, float %tmp305, float %tmp307)327 %tmp309 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 177328 %tmp310 = load float, ptr addrspace(3) %tmp309, align 4329 %tmp311 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 178330 %tmp312 = load float, ptr addrspace(3) %tmp311, align 4331 %tmp313 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 179332 %tmp314 = load float, ptr addrspace(3) %tmp313, align 4333 %tmp315 = tail call float @llvm.fmuladd.f32(float %tmp310, float %tmp312, float %tmp314)334 %tmp316 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 181335 %tmp317 = load float, ptr addrspace(3) %tmp316, align 4336 %tmp318 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 182337 %tmp319 = load float, ptr addrspace(3) %tmp318, align 4338 %tmp320 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 183339 %tmp321 = load float, ptr addrspace(3) %tmp320, align 4340 %tmp322 = tail call float @llvm.fmuladd.f32(float %tmp317, float %tmp319, float %tmp321)341 %tmp323 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 185342 %tmp324 = load float, ptr addrspace(3) %tmp323, align 4343 %tmp325 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 186344 %tmp326 = load float, ptr addrspace(3) %tmp325, align 4345 %tmp327 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 187346 %tmp328 = load float, ptr addrspace(3) %tmp327, align 4347 %tmp329 = tail call float @llvm.fmuladd.f32(float %tmp324, float %tmp326, float %tmp328)348 %tmp330 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 189349 %tmp331 = load float, ptr addrspace(3) %tmp330, align 4350 %tmp332 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 190351 %tmp333 = load float, ptr addrspace(3) %tmp332, align 4352 %tmp334 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 191353 %tmp335 = load float, ptr addrspace(3) %tmp334, align 4354 %tmp336 = tail call float @llvm.fmuladd.f32(float %tmp331, float %tmp333, float %tmp335)355 %tmp337 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 193356 %tmp338 = load float, ptr addrspace(3) %tmp337, align 4357 %tmp339 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 194358 %tmp340 = load float, ptr addrspace(3) %tmp339, align 4359 %tmp341 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 195360 %tmp342 = load float, ptr addrspace(3) %tmp341, align 4361 %tmp343 = tail call float @llvm.fmuladd.f32(float %tmp338, float %tmp340, float %tmp342)362 %tmp344 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 197363 %tmp345 = load float, ptr addrspace(3) %tmp344, align 4364 %tmp346 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 198365 %tmp347 = load float, ptr addrspace(3) %tmp346, align 4366 %tmp348 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 199367 %tmp349 = load float, ptr addrspace(3) %tmp348, align 4368 %tmp350 = tail call float @llvm.fmuladd.f32(float %tmp345, float %tmp347, float %tmp349)369 %tmp351 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 201370 %tmp352 = load float, ptr addrspace(3) %tmp351, align 4371 %tmp353 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 202372 %tmp354 = load float, ptr addrspace(3) %tmp353, align 4373 %tmp355 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 203374 %tmp356 = load float, ptr addrspace(3) %tmp355, align 4375 %tmp357 = tail call float @llvm.fmuladd.f32(float %tmp352, float %tmp354, float %tmp356)376 %tmp358 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 205377 %tmp359 = load float, ptr addrspace(3) %tmp358, align 4378 %tmp360 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 206379 %tmp361 = load float, ptr addrspace(3) %tmp360, align 4380 %tmp362 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 207381 %tmp363 = load float, ptr addrspace(3) %tmp362, align 4382 %tmp364 = tail call float @llvm.fmuladd.f32(float %tmp359, float %tmp361, float %tmp363)383 %tmp365 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 209384 %tmp366 = load float, ptr addrspace(3) %tmp365, align 4385 %tmp367 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 210386 %tmp368 = load float, ptr addrspace(3) %tmp367, align 4387 %tmp369 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 211388 %tmp370 = load float, ptr addrspace(3) %tmp369, align 4389 %tmp371 = tail call float @llvm.fmuladd.f32(float %tmp366, float %tmp368, float %tmp370)390 %tmp372 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 213391 %tmp373 = load float, ptr addrspace(3) %tmp372, align 4392 %tmp374 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 214393 %tmp375 = load float, ptr addrspace(3) %tmp374, align 4394 %tmp376 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 215395 %tmp377 = load float, ptr addrspace(3) %tmp376, align 4396 %tmp378 = tail call float @llvm.fmuladd.f32(float %tmp373, float %tmp375, float %tmp377)397 %tmp379 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 217398 %tmp380 = load float, ptr addrspace(3) %tmp379, align 4399 %tmp381 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 218400 %tmp382 = load float, ptr addrspace(3) %tmp381, align 4401 %tmp383 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 219402 %tmp384 = load float, ptr addrspace(3) %tmp383, align 4403 %tmp385 = tail call float @llvm.fmuladd.f32(float %tmp380, float %tmp382, float %tmp384)404 %tmp386 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 221405 %tmp387 = load float, ptr addrspace(3) %tmp386, align 4406 %tmp388 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 222407 %tmp389 = load float, ptr addrspace(3) %tmp388, align 4408 %tmp390 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 223409 %tmp391 = load float, ptr addrspace(3) %tmp390, align 4410 %tmp392 = tail call float @llvm.fmuladd.f32(float %tmp387, float %tmp389, float %tmp391)411 %tmp393 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 225412 %tmp394 = load float, ptr addrspace(3) %tmp393, align 4413 %tmp395 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 226414 %tmp396 = load float, ptr addrspace(3) %tmp395, align 4415 %tmp397 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 227416 %tmp398 = load float, ptr addrspace(3) %tmp397, align 4417 %tmp399 = tail call float @llvm.fmuladd.f32(float %tmp394, float %tmp396, float %tmp398)418 %tmp400 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 229419 %tmp401 = load float, ptr addrspace(3) %tmp400, align 4420 %tmp402 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 230421 %tmp403 = load float, ptr addrspace(3) %tmp402, align 4422 %tmp404 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 231423 %tmp405 = load float, ptr addrspace(3) %tmp404, align 4424 %tmp406 = tail call float @llvm.fmuladd.f32(float %tmp401, float %tmp403, float %tmp405)425 %tmp407 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 233426 %tmp408 = load float, ptr addrspace(3) %tmp407, align 4427 %tmp409 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 234428 %tmp410 = load float, ptr addrspace(3) %tmp409, align 4429 %tmp411 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 235430 %tmp412 = load float, ptr addrspace(3) %tmp411, align 4431 %tmp413 = tail call float @llvm.fmuladd.f32(float %tmp408, float %tmp410, float %tmp412)432 %tmp414 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 237433 %tmp415 = load float, ptr addrspace(3) %tmp414, align 4434 %tmp416 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 238435 %tmp417 = load float, ptr addrspace(3) %tmp416, align 4436 %tmp418 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 239437 %tmp419 = load float, ptr addrspace(3) %tmp418, align 4438 %tmp420 = tail call float @llvm.fmuladd.f32(float %tmp415, float %tmp417, float %tmp419)439 %tmp421 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 241440 %tmp422 = load float, ptr addrspace(3) %tmp421, align 4441 %tmp423 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 242442 %tmp424 = load float, ptr addrspace(3) %tmp423, align 4443 %tmp425 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 243444 %tmp426 = load float, ptr addrspace(3) %tmp425, align 4445 %tmp427 = tail call float @llvm.fmuladd.f32(float %tmp422, float %tmp424, float %tmp426)446 %tmp428 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 245447 %tmp429 = load float, ptr addrspace(3) %tmp428, align 4448 %tmp430 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 246449 %tmp431 = load float, ptr addrspace(3) %tmp430, align 4450 %tmp432 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 247451 %tmp433 = load float, ptr addrspace(3) %tmp432, align 4452 %tmp434 = tail call float @llvm.fmuladd.f32(float %tmp429, float %tmp431, float %tmp433)453 %tmp435 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 249454 %tmp436 = load float, ptr addrspace(3) %tmp435, align 4455 %tmp437 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 250456 %tmp438 = load float, ptr addrspace(3) %tmp437, align 4457 %tmp439 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 251458 %tmp440 = load float, ptr addrspace(3) %tmp439, align 4459 %tmp441 = tail call float @llvm.fmuladd.f32(float %tmp436, float %tmp438, float %tmp440)460 %tmp442 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 253461 %tmp443 = load float, ptr addrspace(3) %tmp442, align 4462 %tmp444 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 254463 %tmp445 = load float, ptr addrspace(3) %tmp444, align 4464 %tmp446 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 255465 %tmp447 = load float, ptr addrspace(3) %tmp446, align 4466 %tmp448 = tail call float @llvm.fmuladd.f32(float %tmp443, float %tmp445, float %tmp447)467 store float %tmp7, ptr addrspace(1) %arg1, align 4468 %tmp449 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 1469 store float %tmp14, ptr addrspace(1) %tmp449, align 4470 %tmp450 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 2471 store float %tmp21, ptr addrspace(1) %tmp450, align 4472 %tmp451 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 3473 store float %tmp28, ptr addrspace(1) %tmp451, align 4474 %tmp452 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 4475 store float %tmp35, ptr addrspace(1) %tmp452, align 4476 %tmp453 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 5477 store float %tmp42, ptr addrspace(1) %tmp453, align 4478 %tmp454 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 6479 store float %tmp49, ptr addrspace(1) %tmp454, align 4480 %tmp455 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 7481 store float %tmp56, ptr addrspace(1) %tmp455, align 4482 %tmp456 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 8483 store float %tmp63, ptr addrspace(1) %tmp456, align 4484 %tmp457 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 9485 store float %tmp70, ptr addrspace(1) %tmp457, align 4486 %tmp458 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 10487 store float %tmp77, ptr addrspace(1) %tmp458, align 4488 %tmp459 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 11489 store float %tmp84, ptr addrspace(1) %tmp459, align 4490 %tmp460 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 12491 store float %tmp91, ptr addrspace(1) %tmp460, align 4492 %tmp461 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 13493 store float %tmp98, ptr addrspace(1) %tmp461, align 4494 %tmp462 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 14495 store float %tmp105, ptr addrspace(1) %tmp462, align 4496 %tmp463 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 15497 store float %tmp112, ptr addrspace(1) %tmp463, align 4498 %tmp464 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 16499 store float %tmp119, ptr addrspace(1) %tmp464, align 4500 %tmp465 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 17501 store float %tmp126, ptr addrspace(1) %tmp465, align 4502 %tmp466 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 18503 store float %tmp133, ptr addrspace(1) %tmp466, align 4504 %tmp467 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 19505 store float %tmp140, ptr addrspace(1) %tmp467, align 4506 %tmp468 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 20507 store float %tmp147, ptr addrspace(1) %tmp468, align 4508 %tmp469 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 21509 store float %tmp154, ptr addrspace(1) %tmp469, align 4510 %tmp470 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 22511 store float %tmp161, ptr addrspace(1) %tmp470, align 4512 %tmp471 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 23513 store float %tmp168, ptr addrspace(1) %tmp471, align 4514 %tmp472 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 24515 store float %tmp175, ptr addrspace(1) %tmp472, align 4516 %tmp473 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 25517 store float %tmp182, ptr addrspace(1) %tmp473, align 4518 %tmp474 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 26519 store float %tmp189, ptr addrspace(1) %tmp474, align 4520 %tmp475 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 27521 store float %tmp196, ptr addrspace(1) %tmp475, align 4522 %tmp476 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 28523 store float %tmp203, ptr addrspace(1) %tmp476, align 4524 %tmp477 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 29525 store float %tmp210, ptr addrspace(1) %tmp477, align 4526 %tmp478 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 30527 store float %tmp217, ptr addrspace(1) %tmp478, align 4528 %tmp479 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 31529 store float %tmp224, ptr addrspace(1) %tmp479, align 4530 %tmp480 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 32531 store float %tmp231, ptr addrspace(1) %tmp480, align 4532 %tmp481 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 33533 store float %tmp238, ptr addrspace(1) %tmp481, align 4534 %tmp482 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 34535 store float %tmp245, ptr addrspace(1) %tmp482, align 4536 %tmp483 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 35537 store float %tmp252, ptr addrspace(1) %tmp483, align 4538 %tmp484 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 36539 store float %tmp259, ptr addrspace(1) %tmp484, align 4540 %tmp485 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 37541 store float %tmp266, ptr addrspace(1) %tmp485, align 4542 %tmp486 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 38543 store float %tmp273, ptr addrspace(1) %tmp486, align 4544 %tmp487 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 39545 store float %tmp280, ptr addrspace(1) %tmp487, align 4546 %tmp488 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 40547 store float %tmp287, ptr addrspace(1) %tmp488, align 4548 %tmp489 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 41549 store float %tmp294, ptr addrspace(1) %tmp489, align 4550 %tmp490 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 42551 store float %tmp301, ptr addrspace(1) %tmp490, align 4552 %tmp491 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 43553 store float %tmp308, ptr addrspace(1) %tmp491, align 4554 %tmp492 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 44555 store float %tmp315, ptr addrspace(1) %tmp492, align 4556 %tmp493 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 45557 store float %tmp322, ptr addrspace(1) %tmp493, align 4558 %tmp494 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 46559 store float %tmp329, ptr addrspace(1) %tmp494, align 4560 %tmp495 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 47561 store float %tmp336, ptr addrspace(1) %tmp495, align 4562 %tmp496 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 48563 store float %tmp343, ptr addrspace(1) %tmp496, align 4564 %tmp497 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 49565 store float %tmp350, ptr addrspace(1) %tmp497, align 4566 %tmp498 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 50567 store float %tmp357, ptr addrspace(1) %tmp498, align 4568 %tmp499 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 51569 store float %tmp364, ptr addrspace(1) %tmp499, align 4570 %tmp500 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 52571 store float %tmp371, ptr addrspace(1) %tmp500, align 4572 %tmp501 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 53573 store float %tmp378, ptr addrspace(1) %tmp501, align 4574 %tmp502 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 54575 store float %tmp385, ptr addrspace(1) %tmp502, align 4576 %tmp503 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 55577 store float %tmp392, ptr addrspace(1) %tmp503, align 4578 %tmp504 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 56579 store float %tmp399, ptr addrspace(1) %tmp504, align 4580 %tmp505 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 57581 store float %tmp406, ptr addrspace(1) %tmp505, align 4582 %tmp506 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 58583 store float %tmp413, ptr addrspace(1) %tmp506, align 4584 %tmp507 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 59585 store float %tmp420, ptr addrspace(1) %tmp507, align 4586 %tmp508 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 60587 store float %tmp427, ptr addrspace(1) %tmp508, align 4588 %tmp509 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 61589 store float %tmp434, ptr addrspace(1) %tmp509, align 4590 %tmp510 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 62591 store float %tmp441, ptr addrspace(1) %tmp510, align 4592 %tmp511 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 63593 store float %tmp448, ptr addrspace(1) %tmp511, align 4594 ret void595}596 597; Function Attrs: nounwind readnone598declare float @llvm.fmuladd.f32(float, float, float) #0599 600attributes #0 = { nounwind readnone }601attributes #1 = { "amdgpu-waves-per-eu"="1,1" "amdgpu-flat-work-group-size"="1,256" }602