608 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=gfx906 < %s | FileCheck --check-prefix=OCC %s2; RUN: llc -mtriple=amdgcn -mcpu=gfx906 -amdgpu-use-amdgpu-trackers=1 < %s | FileCheck --check-prefix=OCC-GCNTRACKER %s3; RUN: llc -mtriple=amdgcn -mcpu=gfx906 -amdgpu-schedule-relaxed-occupancy=true < %s | FileCheck --check-prefix=RELAX %s4; RUN: llc -mtriple=amdgcn -mcpu=gfx906 -amdgpu-use-amdgpu-trackers=1 -amdgpu-schedule-relaxed-occupancy=true < %s | FileCheck --check-prefix=RELAX-GCNTRACKER %s5 6 7; Using -amgpu-schedule-relaxed-occupancy allows scheduler to produce better ILP by further relaxing occupancy target8 9; CHECK-LABEL: {{^}}load_fma_store:10; OCC: NumVgprs: 2411; OCC-GCNTRACKER: NumVgprs: 2612; RELAX: NumVgprs: 6413; RELAX-GCNTRACKER: NumVgprs: 6014; OCC: NumVGPRsForWavesPerEU: 2415; OCC-GCNTRACKER: NumVGPRsForWavesPerEU: 2616; RELAX: NumVGPRsForWavesPerEU: 6417; RELAX-GCNTRACKER: NumVGPRsForWavesPerEU: 6018; OCC: Occupancy: 1019; OCC-GCNTRACKER: Occupancy: 920; RELAX: Occupancy: 421; RELAX-GCNTRACKER: Occupancy: 422 23define amdgpu_kernel void @load_fma_store(ptr addrspace(3) nocapture readonly %arg, ptr addrspace(1) nocapture %arg1) #1 {24bb:25 %tmp = getelementptr inbounds float, ptr addrspace(3) %arg, i32 126 %tmp2 = load float, ptr addrspace(3) %tmp, align 427 %tmp3 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 228 %tmp4 = load float, ptr addrspace(3) %tmp3, align 429 %tmp5 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 330 %tmp6 = load float, ptr addrspace(3) %tmp5, align 431 %tmp7 = tail call float @llvm.fmuladd.f32(float %tmp2, float %tmp4, float %tmp6)32 %tmp8 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 533 %tmp9 = load float, ptr addrspace(3) %tmp8, align 434 %tmp10 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 635 %tmp11 = load float, ptr addrspace(3) %tmp10, align 436 %tmp12 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 737 %tmp13 = load float, ptr addrspace(3) %tmp12, align 438 %tmp14 = tail call float @llvm.fmuladd.f32(float %tmp9, float %tmp11, float %tmp13)39 %tmp15 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 940 %tmp16 = load float, ptr addrspace(3) %tmp15, align 441 %tmp17 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1042 %tmp18 = load float, ptr addrspace(3) %tmp17, align 443 %tmp19 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1144 %tmp20 = load float, ptr addrspace(3) %tmp19, align 445 %tmp21 = tail call float @llvm.fmuladd.f32(float %tmp16, float %tmp18, float %tmp20)46 %tmp22 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1347 %tmp23 = load float, ptr addrspace(3) %tmp22, align 448 %tmp24 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1449 %tmp25 = load float, ptr addrspace(3) %tmp24, align 450 %tmp26 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1551 %tmp27 = load float, ptr addrspace(3) %tmp26, align 452 %tmp28 = tail call float @llvm.fmuladd.f32(float %tmp23, float %tmp25, float %tmp27)53 %tmp29 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1754 %tmp30 = load float, ptr addrspace(3) %tmp29, align 455 %tmp31 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1856 %tmp32 = load float, ptr addrspace(3) %tmp31, align 457 %tmp33 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1958 %tmp34 = load float, ptr addrspace(3) %tmp33, align 459 %tmp35 = tail call float @llvm.fmuladd.f32(float %tmp30, float %tmp32, float %tmp34)60 %tmp36 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2161 %tmp37 = load float, ptr addrspace(3) %tmp36, align 462 %tmp38 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2263 %tmp39 = load float, ptr addrspace(3) %tmp38, align 464 %tmp40 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2365 %tmp41 = load float, ptr addrspace(3) %tmp40, align 466 %tmp42 = tail call float @llvm.fmuladd.f32(float %tmp37, float %tmp39, float %tmp41)67 %tmp43 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2568 %tmp44 = load float, ptr addrspace(3) %tmp43, align 469 %tmp45 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2670 %tmp46 = load float, ptr addrspace(3) %tmp45, align 471 %tmp47 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2772 %tmp48 = load float, ptr addrspace(3) %tmp47, align 473 %tmp49 = tail call float @llvm.fmuladd.f32(float %tmp44, float %tmp46, float %tmp48)74 %tmp50 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2975 %tmp51 = load float, ptr addrspace(3) %tmp50, align 476 %tmp52 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3077 %tmp53 = load float, ptr addrspace(3) %tmp52, align 478 %tmp54 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3179 %tmp55 = load float, ptr addrspace(3) %tmp54, align 480 %tmp56 = tail call float @llvm.fmuladd.f32(float %tmp51, float %tmp53, float %tmp55)81 %tmp57 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3382 %tmp58 = load float, ptr addrspace(3) %tmp57, align 483 %tmp59 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3484 %tmp60 = load float, ptr addrspace(3) %tmp59, align 485 %tmp61 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3586 %tmp62 = load float, ptr addrspace(3) %tmp61, align 487 %tmp63 = tail call float @llvm.fmuladd.f32(float %tmp58, float %tmp60, float %tmp62)88 %tmp64 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3789 %tmp65 = load float, ptr addrspace(3) %tmp64, align 490 %tmp66 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3891 %tmp67 = load float, ptr addrspace(3) %tmp66, align 492 %tmp68 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3993 %tmp69 = load float, ptr addrspace(3) %tmp68, align 494 %tmp70 = tail call float @llvm.fmuladd.f32(float %tmp65, float %tmp67, float %tmp69)95 %tmp71 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4196 %tmp72 = load float, ptr addrspace(3) %tmp71, align 497 %tmp73 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 4298 %tmp74 = load float, ptr addrspace(3) %tmp73, align 499 %tmp75 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 43100 %tmp76 = load float, ptr addrspace(3) %tmp75, align 4101 %tmp77 = tail call float @llvm.fmuladd.f32(float %tmp72, float %tmp74, float %tmp76)102 %tmp78 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 45103 %tmp79 = load float, ptr addrspace(3) %tmp78, align 4104 %tmp80 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 46105 %tmp81 = load float, ptr addrspace(3) %tmp80, align 4106 %tmp82 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 47107 %tmp83 = load float, ptr addrspace(3) %tmp82, align 4108 %tmp84 = tail call float @llvm.fmuladd.f32(float %tmp79, float %tmp81, float %tmp83)109 %tmp85 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 49110 %tmp86 = load float, ptr addrspace(3) %tmp85, align 4111 %tmp87 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 50112 %tmp88 = load float, ptr addrspace(3) %tmp87, align 4113 %tmp89 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 51114 %tmp90 = load float, ptr addrspace(3) %tmp89, align 4115 %tmp91 = tail call float @llvm.fmuladd.f32(float %tmp86, float %tmp88, float %tmp90)116 %tmp92 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 53117 %tmp93 = load float, ptr addrspace(3) %tmp92, align 4118 %tmp94 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 54119 %tmp95 = load float, ptr addrspace(3) %tmp94, align 4120 %tmp96 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 55121 %tmp97 = load float, ptr addrspace(3) %tmp96, align 4122 %tmp98 = tail call float @llvm.fmuladd.f32(float %tmp93, float %tmp95, float %tmp97)123 %tmp99 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 57124 %tmp100 = load float, ptr addrspace(3) %tmp99, align 4125 %tmp101 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 58126 %tmp102 = load float, ptr addrspace(3) %tmp101, align 4127 %tmp103 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 59128 %tmp104 = load float, ptr addrspace(3) %tmp103, align 4129 %tmp105 = tail call float @llvm.fmuladd.f32(float %tmp100, float %tmp102, float %tmp104)130 %tmp106 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 61131 %tmp107 = load float, ptr addrspace(3) %tmp106, align 4132 %tmp108 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 62133 %tmp109 = load float, ptr addrspace(3) %tmp108, align 4134 %tmp110 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 63135 %tmp111 = load float, ptr addrspace(3) %tmp110, align 4136 %tmp112 = tail call float @llvm.fmuladd.f32(float %tmp107, float %tmp109, float %tmp111)137 %tmp113 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 65138 %tmp114 = load float, ptr addrspace(3) %tmp113, align 4139 %tmp115 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 66140 %tmp116 = load float, ptr addrspace(3) %tmp115, align 4141 %tmp117 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 67142 %tmp118 = load float, ptr addrspace(3) %tmp117, align 4143 %tmp119 = tail call float @llvm.fmuladd.f32(float %tmp114, float %tmp116, float %tmp118)144 %tmp120 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 69145 %tmp121 = load float, ptr addrspace(3) %tmp120, align 4146 %tmp122 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 70147 %tmp123 = load float, ptr addrspace(3) %tmp122, align 4148 %tmp124 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 71149 %tmp125 = load float, ptr addrspace(3) %tmp124, align 4150 %tmp126 = tail call float @llvm.fmuladd.f32(float %tmp121, float %tmp123, float %tmp125)151 %tmp127 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 73152 %tmp128 = load float, ptr addrspace(3) %tmp127, align 4153 %tmp129 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 74154 %tmp130 = load float, ptr addrspace(3) %tmp129, align 4155 %tmp131 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 75156 %tmp132 = load float, ptr addrspace(3) %tmp131, align 4157 %tmp133 = tail call float @llvm.fmuladd.f32(float %tmp128, float %tmp130, float %tmp132)158 %tmp134 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 77159 %tmp135 = load float, ptr addrspace(3) %tmp134, align 4160 %tmp136 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 78161 %tmp137 = load float, ptr addrspace(3) %tmp136, align 4162 %tmp138 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 79163 %tmp139 = load float, ptr addrspace(3) %tmp138, align 4164 %tmp140 = tail call float @llvm.fmuladd.f32(float %tmp135, float %tmp137, float %tmp139)165 %tmp141 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 81166 %tmp142 = load float, ptr addrspace(3) %tmp141, align 4167 %tmp143 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 82168 %tmp144 = load float, ptr addrspace(3) %tmp143, align 4169 %tmp145 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 83170 %tmp146 = load float, ptr addrspace(3) %tmp145, align 4171 %tmp147 = tail call float @llvm.fmuladd.f32(float %tmp142, float %tmp144, float %tmp146)172 %tmp148 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 85173 %tmp149 = load float, ptr addrspace(3) %tmp148, align 4174 %tmp150 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 86175 %tmp151 = load float, ptr addrspace(3) %tmp150, align 4176 %tmp152 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 87177 %tmp153 = load float, ptr addrspace(3) %tmp152, align 4178 %tmp154 = tail call float @llvm.fmuladd.f32(float %tmp149, float %tmp151, float %tmp153)179 %tmp155 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 89180 %tmp156 = load float, ptr addrspace(3) %tmp155, align 4181 %tmp157 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 90182 %tmp158 = load float, ptr addrspace(3) %tmp157, align 4183 %tmp159 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 91184 %tmp160 = load float, ptr addrspace(3) %tmp159, align 4185 %tmp161 = tail call float @llvm.fmuladd.f32(float %tmp156, float %tmp158, float %tmp160)186 %tmp162 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 93187 %tmp163 = load float, ptr addrspace(3) %tmp162, align 4188 %tmp164 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 94189 %tmp165 = load float, ptr addrspace(3) %tmp164, align 4190 %tmp166 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 95191 %tmp167 = load float, ptr addrspace(3) %tmp166, align 4192 %tmp168 = tail call float @llvm.fmuladd.f32(float %tmp163, float %tmp165, float %tmp167)193 %tmp169 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 97194 %tmp170 = load float, ptr addrspace(3) %tmp169, align 4195 %tmp171 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 98196 %tmp172 = load float, ptr addrspace(3) %tmp171, align 4197 %tmp173 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 99198 %tmp174 = load float, ptr addrspace(3) %tmp173, align 4199 %tmp175 = tail call float @llvm.fmuladd.f32(float %tmp170, float %tmp172, float %tmp174)200 %tmp176 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 101201 %tmp177 = load float, ptr addrspace(3) %tmp176, align 4202 %tmp178 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 102203 %tmp179 = load float, ptr addrspace(3) %tmp178, align 4204 %tmp180 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 103205 %tmp181 = load float, ptr addrspace(3) %tmp180, align 4206 %tmp182 = tail call float @llvm.fmuladd.f32(float %tmp177, float %tmp179, float %tmp181)207 %tmp183 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 105208 %tmp184 = load float, ptr addrspace(3) %tmp183, align 4209 %tmp185 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 106210 %tmp186 = load float, ptr addrspace(3) %tmp185, align 4211 %tmp187 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 107212 %tmp188 = load float, ptr addrspace(3) %tmp187, align 4213 %tmp189 = tail call float @llvm.fmuladd.f32(float %tmp184, float %tmp186, float %tmp188)214 %tmp190 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 109215 %tmp191 = load float, ptr addrspace(3) %tmp190, align 4216 %tmp192 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 110217 %tmp193 = load float, ptr addrspace(3) %tmp192, align 4218 %tmp194 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 111219 %tmp195 = load float, ptr addrspace(3) %tmp194, align 4220 %tmp196 = tail call float @llvm.fmuladd.f32(float %tmp191, float %tmp193, float %tmp195)221 %tmp197 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 113222 %tmp198 = load float, ptr addrspace(3) %tmp197, align 4223 %tmp199 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 114224 %tmp200 = load float, ptr addrspace(3) %tmp199, align 4225 %tmp201 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 115226 %tmp202 = load float, ptr addrspace(3) %tmp201, align 4227 %tmp203 = tail call float @llvm.fmuladd.f32(float %tmp198, float %tmp200, float %tmp202)228 %tmp204 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 117229 %tmp205 = load float, ptr addrspace(3) %tmp204, align 4230 %tmp206 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 118231 %tmp207 = load float, ptr addrspace(3) %tmp206, align 4232 %tmp208 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 119233 %tmp209 = load float, ptr addrspace(3) %tmp208, align 4234 %tmp210 = tail call float @llvm.fmuladd.f32(float %tmp205, float %tmp207, float %tmp209)235 %tmp211 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 121236 %tmp212 = load float, ptr addrspace(3) %tmp211, align 4237 %tmp213 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 122238 %tmp214 = load float, ptr addrspace(3) %tmp213, align 4239 %tmp215 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 123240 %tmp216 = load float, ptr addrspace(3) %tmp215, align 4241 %tmp217 = tail call float @llvm.fmuladd.f32(float %tmp212, float %tmp214, float %tmp216)242 %tmp218 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 125243 %tmp219 = load float, ptr addrspace(3) %tmp218, align 4244 %tmp220 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 126245 %tmp221 = load float, ptr addrspace(3) %tmp220, align 4246 %tmp222 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 127247 %tmp223 = load float, ptr addrspace(3) %tmp222, align 4248 %tmp224 = tail call float @llvm.fmuladd.f32(float %tmp219, float %tmp221, float %tmp223)249 %tmp225 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 129250 %tmp226 = load float, ptr addrspace(3) %tmp225, align 4251 %tmp227 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 130252 %tmp228 = load float, ptr addrspace(3) %tmp227, align 4253 %tmp229 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 131254 %tmp230 = load float, ptr addrspace(3) %tmp229, align 4255 %tmp231 = tail call float @llvm.fmuladd.f32(float %tmp226, float %tmp228, float %tmp230)256 %tmp232 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 133257 %tmp233 = load float, ptr addrspace(3) %tmp232, align 4258 %tmp234 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 134259 %tmp235 = load float, ptr addrspace(3) %tmp234, align 4260 %tmp236 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 135261 %tmp237 = load float, ptr addrspace(3) %tmp236, align 4262 %tmp238 = tail call float @llvm.fmuladd.f32(float %tmp233, float %tmp235, float %tmp237)263 %tmp239 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 137264 %tmp240 = load float, ptr addrspace(3) %tmp239, align 4265 %tmp241 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 138266 %tmp242 = load float, ptr addrspace(3) %tmp241, align 4267 %tmp243 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 139268 %tmp244 = load float, ptr addrspace(3) %tmp243, align 4269 %tmp245 = tail call float @llvm.fmuladd.f32(float %tmp240, float %tmp242, float %tmp244)270 %tmp246 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 141271 %tmp247 = load float, ptr addrspace(3) %tmp246, align 4272 %tmp248 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 142273 %tmp249 = load float, ptr addrspace(3) %tmp248, align 4274 %tmp250 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 143275 %tmp251 = load float, ptr addrspace(3) %tmp250, align 4276 %tmp252 = tail call float @llvm.fmuladd.f32(float %tmp247, float %tmp249, float %tmp251)277 %tmp253 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 145278 %tmp254 = load float, ptr addrspace(3) %tmp253, align 4279 %tmp255 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 146280 %tmp256 = load float, ptr addrspace(3) %tmp255, align 4281 %tmp257 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 147282 %tmp258 = load float, ptr addrspace(3) %tmp257, align 4283 %tmp259 = tail call float @llvm.fmuladd.f32(float %tmp254, float %tmp256, float %tmp258)284 %tmp260 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 149285 %tmp261 = load float, ptr addrspace(3) %tmp260, align 4286 %tmp262 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 150287 %tmp263 = load float, ptr addrspace(3) %tmp262, align 4288 %tmp264 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 151289 %tmp265 = load float, ptr addrspace(3) %tmp264, align 4290 %tmp266 = tail call float @llvm.fmuladd.f32(float %tmp261, float %tmp263, float %tmp265)291 %tmp267 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 153292 %tmp268 = load float, ptr addrspace(3) %tmp267, align 4293 %tmp269 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 154294 %tmp270 = load float, ptr addrspace(3) %tmp269, align 4295 %tmp271 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 155296 %tmp272 = load float, ptr addrspace(3) %tmp271, align 4297 %tmp273 = tail call float @llvm.fmuladd.f32(float %tmp268, float %tmp270, float %tmp272)298 %tmp274 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 157299 %tmp275 = load float, ptr addrspace(3) %tmp274, align 4300 %tmp276 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 158301 %tmp277 = load float, ptr addrspace(3) %tmp276, align 4302 %tmp278 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 159303 %tmp279 = load float, ptr addrspace(3) %tmp278, align 4304 %tmp280 = tail call float @llvm.fmuladd.f32(float %tmp275, float %tmp277, float %tmp279)305 %tmp281 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 161306 %tmp282 = load float, ptr addrspace(3) %tmp281, align 4307 %tmp283 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 162308 %tmp284 = load float, ptr addrspace(3) %tmp283, align 4309 %tmp285 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 163310 %tmp286 = load float, ptr addrspace(3) %tmp285, align 4311 %tmp287 = tail call float @llvm.fmuladd.f32(float %tmp282, float %tmp284, float %tmp286)312 %tmp288 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 165313 %tmp289 = load float, ptr addrspace(3) %tmp288, align 4314 %tmp290 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 166315 %tmp291 = load float, ptr addrspace(3) %tmp290, align 4316 %tmp292 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 167317 %tmp293 = load float, ptr addrspace(3) %tmp292, align 4318 %tmp294 = tail call float @llvm.fmuladd.f32(float %tmp289, float %tmp291, float %tmp293)319 %tmp295 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 169320 %tmp296 = load float, ptr addrspace(3) %tmp295, align 4321 %tmp297 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 170322 %tmp298 = load float, ptr addrspace(3) %tmp297, align 4323 %tmp299 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 171324 %tmp300 = load float, ptr addrspace(3) %tmp299, align 4325 %tmp301 = tail call float @llvm.fmuladd.f32(float %tmp296, float %tmp298, float %tmp300)326 %tmp302 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 173327 %tmp303 = load float, ptr addrspace(3) %tmp302, align 4328 %tmp304 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 174329 %tmp305 = load float, ptr addrspace(3) %tmp304, align 4330 %tmp306 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 175331 %tmp307 = load float, ptr addrspace(3) %tmp306, align 4332 %tmp308 = tail call float @llvm.fmuladd.f32(float %tmp303, float %tmp305, float %tmp307)333 %tmp309 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 177334 %tmp310 = load float, ptr addrspace(3) %tmp309, align 4335 %tmp311 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 178336 %tmp312 = load float, ptr addrspace(3) %tmp311, align 4337 %tmp313 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 179338 %tmp314 = load float, ptr addrspace(3) %tmp313, align 4339 %tmp315 = tail call float @llvm.fmuladd.f32(float %tmp310, float %tmp312, float %tmp314)340 %tmp316 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 181341 %tmp317 = load float, ptr addrspace(3) %tmp316, align 4342 %tmp318 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 182343 %tmp319 = load float, ptr addrspace(3) %tmp318, align 4344 %tmp320 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 183345 %tmp321 = load float, ptr addrspace(3) %tmp320, align 4346 %tmp322 = tail call float @llvm.fmuladd.f32(float %tmp317, float %tmp319, float %tmp321)347 %tmp323 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 185348 %tmp324 = load float, ptr addrspace(3) %tmp323, align 4349 %tmp325 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 186350 %tmp326 = load float, ptr addrspace(3) %tmp325, align 4351 %tmp327 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 187352 %tmp328 = load float, ptr addrspace(3) %tmp327, align 4353 %tmp329 = tail call float @llvm.fmuladd.f32(float %tmp324, float %tmp326, float %tmp328)354 %tmp330 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 189355 %tmp331 = load float, ptr addrspace(3) %tmp330, align 4356 %tmp332 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 190357 %tmp333 = load float, ptr addrspace(3) %tmp332, align 4358 %tmp334 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 191359 %tmp335 = load float, ptr addrspace(3) %tmp334, align 4360 %tmp336 = tail call float @llvm.fmuladd.f32(float %tmp331, float %tmp333, float %tmp335)361 %tmp337 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 193362 %tmp338 = load float, ptr addrspace(3) %tmp337, align 4363 %tmp339 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 194364 %tmp340 = load float, ptr addrspace(3) %tmp339, align 4365 %tmp341 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 195366 %tmp342 = load float, ptr addrspace(3) %tmp341, align 4367 %tmp343 = tail call float @llvm.fmuladd.f32(float %tmp338, float %tmp340, float %tmp342)368 %tmp344 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 197369 %tmp345 = load float, ptr addrspace(3) %tmp344, align 4370 %tmp346 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 198371 %tmp347 = load float, ptr addrspace(3) %tmp346, align 4372 %tmp348 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 199373 %tmp349 = load float, ptr addrspace(3) %tmp348, align 4374 %tmp350 = tail call float @llvm.fmuladd.f32(float %tmp345, float %tmp347, float %tmp349)375 %tmp351 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 201376 %tmp352 = load float, ptr addrspace(3) %tmp351, align 4377 %tmp353 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 202378 %tmp354 = load float, ptr addrspace(3) %tmp353, align 4379 %tmp355 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 203380 %tmp356 = load float, ptr addrspace(3) %tmp355, align 4381 %tmp357 = tail call float @llvm.fmuladd.f32(float %tmp352, float %tmp354, float %tmp356)382 %tmp358 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 205383 %tmp359 = load float, ptr addrspace(3) %tmp358, align 4384 %tmp360 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 206385 %tmp361 = load float, ptr addrspace(3) %tmp360, align 4386 %tmp362 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 207387 %tmp363 = load float, ptr addrspace(3) %tmp362, align 4388 %tmp364 = tail call float @llvm.fmuladd.f32(float %tmp359, float %tmp361, float %tmp363)389 %tmp365 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 209390 %tmp366 = load float, ptr addrspace(3) %tmp365, align 4391 %tmp367 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 210392 %tmp368 = load float, ptr addrspace(3) %tmp367, align 4393 %tmp369 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 211394 %tmp370 = load float, ptr addrspace(3) %tmp369, align 4395 %tmp371 = tail call float @llvm.fmuladd.f32(float %tmp366, float %tmp368, float %tmp370)396 %tmp372 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 213397 %tmp373 = load float, ptr addrspace(3) %tmp372, align 4398 %tmp374 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 214399 %tmp375 = load float, ptr addrspace(3) %tmp374, align 4400 %tmp376 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 215401 %tmp377 = load float, ptr addrspace(3) %tmp376, align 4402 %tmp378 = tail call float @llvm.fmuladd.f32(float %tmp373, float %tmp375, float %tmp377)403 %tmp379 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 217404 %tmp380 = load float, ptr addrspace(3) %tmp379, align 4405 %tmp381 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 218406 %tmp382 = load float, ptr addrspace(3) %tmp381, align 4407 %tmp383 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 219408 %tmp384 = load float, ptr addrspace(3) %tmp383, align 4409 %tmp385 = tail call float @llvm.fmuladd.f32(float %tmp380, float %tmp382, float %tmp384)410 %tmp386 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 221411 %tmp387 = load float, ptr addrspace(3) %tmp386, align 4412 %tmp388 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 222413 %tmp389 = load float, ptr addrspace(3) %tmp388, align 4414 %tmp390 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 223415 %tmp391 = load float, ptr addrspace(3) %tmp390, align 4416 %tmp392 = tail call float @llvm.fmuladd.f32(float %tmp387, float %tmp389, float %tmp391)417 %tmp393 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 225418 %tmp394 = load float, ptr addrspace(3) %tmp393, align 4419 %tmp395 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 226420 %tmp396 = load float, ptr addrspace(3) %tmp395, align 4421 %tmp397 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 227422 %tmp398 = load float, ptr addrspace(3) %tmp397, align 4423 %tmp399 = tail call float @llvm.fmuladd.f32(float %tmp394, float %tmp396, float %tmp398)424 %tmp400 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 229425 %tmp401 = load float, ptr addrspace(3) %tmp400, align 4426 %tmp402 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 230427 %tmp403 = load float, ptr addrspace(3) %tmp402, align 4428 %tmp404 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 231429 %tmp405 = load float, ptr addrspace(3) %tmp404, align 4430 %tmp406 = tail call float @llvm.fmuladd.f32(float %tmp401, float %tmp403, float %tmp405)431 %tmp407 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 233432 %tmp408 = load float, ptr addrspace(3) %tmp407, align 4433 %tmp409 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 234434 %tmp410 = load float, ptr addrspace(3) %tmp409, align 4435 %tmp411 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 235436 %tmp412 = load float, ptr addrspace(3) %tmp411, align 4437 %tmp413 = tail call float @llvm.fmuladd.f32(float %tmp408, float %tmp410, float %tmp412)438 %tmp414 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 237439 %tmp415 = load float, ptr addrspace(3) %tmp414, align 4440 %tmp416 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 238441 %tmp417 = load float, ptr addrspace(3) %tmp416, align 4442 %tmp418 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 239443 %tmp419 = load float, ptr addrspace(3) %tmp418, align 4444 %tmp420 = tail call float @llvm.fmuladd.f32(float %tmp415, float %tmp417, float %tmp419)445 %tmp421 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 241446 %tmp422 = load float, ptr addrspace(3) %tmp421, align 4447 %tmp423 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 242448 %tmp424 = load float, ptr addrspace(3) %tmp423, align 4449 %tmp425 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 243450 %tmp426 = load float, ptr addrspace(3) %tmp425, align 4451 %tmp427 = tail call float @llvm.fmuladd.f32(float %tmp422, float %tmp424, float %tmp426)452 %tmp428 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 245453 %tmp429 = load float, ptr addrspace(3) %tmp428, align 4454 %tmp430 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 246455 %tmp431 = load float, ptr addrspace(3) %tmp430, align 4456 %tmp432 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 247457 %tmp433 = load float, ptr addrspace(3) %tmp432, align 4458 %tmp434 = tail call float @llvm.fmuladd.f32(float %tmp429, float %tmp431, float %tmp433)459 %tmp435 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 249460 %tmp436 = load float, ptr addrspace(3) %tmp435, align 4461 %tmp437 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 250462 %tmp438 = load float, ptr addrspace(3) %tmp437, align 4463 %tmp439 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 251464 %tmp440 = load float, ptr addrspace(3) %tmp439, align 4465 %tmp441 = tail call float @llvm.fmuladd.f32(float %tmp436, float %tmp438, float %tmp440)466 %tmp442 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 253467 %tmp443 = load float, ptr addrspace(3) %tmp442, align 4468 %tmp444 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 254469 %tmp445 = load float, ptr addrspace(3) %tmp444, align 4470 %tmp446 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 255471 %tmp447 = load float, ptr addrspace(3) %tmp446, align 4472 %tmp448 = tail call float @llvm.fmuladd.f32(float %tmp443, float %tmp445, float %tmp447)473 store float %tmp7, ptr addrspace(1) %arg1, align 4474 %tmp449 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 1475 store float %tmp14, ptr addrspace(1) %tmp449, align 4476 %tmp450 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 2477 store float %tmp21, ptr addrspace(1) %tmp450, align 4478 %tmp451 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 3479 store float %tmp28, ptr addrspace(1) %tmp451, align 4480 %tmp452 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 4481 store float %tmp35, ptr addrspace(1) %tmp452, align 4482 %tmp453 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 5483 store float %tmp42, ptr addrspace(1) %tmp453, align 4484 %tmp454 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 6485 store float %tmp49, ptr addrspace(1) %tmp454, align 4486 %tmp455 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 7487 store float %tmp56, ptr addrspace(1) %tmp455, align 4488 %tmp456 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 8489 store float %tmp63, ptr addrspace(1) %tmp456, align 4490 %tmp457 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 9491 store float %tmp70, ptr addrspace(1) %tmp457, align 4492 %tmp458 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 10493 store float %tmp77, ptr addrspace(1) %tmp458, align 4494 %tmp459 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 11495 store float %tmp84, ptr addrspace(1) %tmp459, align 4496 %tmp460 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 12497 store float %tmp91, ptr addrspace(1) %tmp460, align 4498 %tmp461 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 13499 store float %tmp98, ptr addrspace(1) %tmp461, align 4500 %tmp462 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 14501 store float %tmp105, ptr addrspace(1) %tmp462, align 4502 %tmp463 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 15503 store float %tmp112, ptr addrspace(1) %tmp463, align 4504 %tmp464 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 16505 store float %tmp119, ptr addrspace(1) %tmp464, align 4506 %tmp465 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 17507 store float %tmp126, ptr addrspace(1) %tmp465, align 4508 %tmp466 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 18509 store float %tmp133, ptr addrspace(1) %tmp466, align 4510 %tmp467 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 19511 store float %tmp140, ptr addrspace(1) %tmp467, align 4512 %tmp468 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 20513 store float %tmp147, ptr addrspace(1) %tmp468, align 4514 %tmp469 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 21515 store float %tmp154, ptr addrspace(1) %tmp469, align 4516 %tmp470 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 22517 store float %tmp161, ptr addrspace(1) %tmp470, align 4518 %tmp471 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 23519 store float %tmp168, ptr addrspace(1) %tmp471, align 4520 %tmp472 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 24521 store float %tmp175, ptr addrspace(1) %tmp472, align 4522 %tmp473 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 25523 store float %tmp182, ptr addrspace(1) %tmp473, align 4524 %tmp474 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 26525 store float %tmp189, ptr addrspace(1) %tmp474, align 4526 %tmp475 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 27527 store float %tmp196, ptr addrspace(1) %tmp475, align 4528 %tmp476 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 28529 store float %tmp203, ptr addrspace(1) %tmp476, align 4530 %tmp477 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 29531 store float %tmp210, ptr addrspace(1) %tmp477, align 4532 %tmp478 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 30533 store float %tmp217, ptr addrspace(1) %tmp478, align 4534 %tmp479 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 31535 store float %tmp224, ptr addrspace(1) %tmp479, align 4536 %tmp480 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 32537 store float %tmp231, ptr addrspace(1) %tmp480, align 4538 %tmp481 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 33539 store float %tmp238, ptr addrspace(1) %tmp481, align 4540 %tmp482 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 34541 store float %tmp245, ptr addrspace(1) %tmp482, align 4542 %tmp483 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 35543 store float %tmp252, ptr addrspace(1) %tmp483, align 4544 %tmp484 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 36545 store float %tmp259, ptr addrspace(1) %tmp484, align 4546 %tmp485 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 37547 store float %tmp266, ptr addrspace(1) %tmp485, align 4548 %tmp486 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 38549 store float %tmp273, ptr addrspace(1) %tmp486, align 4550 %tmp487 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 39551 store float %tmp280, ptr addrspace(1) %tmp487, align 4552 %tmp488 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 40553 store float %tmp287, ptr addrspace(1) %tmp488, align 4554 %tmp489 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 41555 store float %tmp294, ptr addrspace(1) %tmp489, align 4556 %tmp490 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 42557 store float %tmp301, ptr addrspace(1) %tmp490, align 4558 %tmp491 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 43559 store float %tmp308, ptr addrspace(1) %tmp491, align 4560 %tmp492 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 44561 store float %tmp315, ptr addrspace(1) %tmp492, align 4562 %tmp493 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 45563 store float %tmp322, ptr addrspace(1) %tmp493, align 4564 %tmp494 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 46565 store float %tmp329, ptr addrspace(1) %tmp494, align 4566 %tmp495 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 47567 store float %tmp336, ptr addrspace(1) %tmp495, align 4568 %tmp496 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 48569 store float %tmp343, ptr addrspace(1) %tmp496, align 4570 %tmp497 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 49571 store float %tmp350, ptr addrspace(1) %tmp497, align 4572 %tmp498 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 50573 store float %tmp357, ptr addrspace(1) %tmp498, align 4574 %tmp499 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 51575 store float %tmp364, ptr addrspace(1) %tmp499, align 4576 %tmp500 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 52577 store float %tmp371, ptr addrspace(1) %tmp500, align 4578 %tmp501 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 53579 store float %tmp378, ptr addrspace(1) %tmp501, align 4580 %tmp502 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 54581 store float %tmp385, ptr addrspace(1) %tmp502, align 4582 %tmp503 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 55583 store float %tmp392, ptr addrspace(1) %tmp503, align 4584 %tmp504 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 56585 store float %tmp399, ptr addrspace(1) %tmp504, align 4586 %tmp505 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 57587 store float %tmp406, ptr addrspace(1) %tmp505, align 4588 %tmp506 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 58589 store float %tmp413, ptr addrspace(1) %tmp506, align 4590 %tmp507 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 59591 store float %tmp420, ptr addrspace(1) %tmp507, align 4592 %tmp508 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 60593 store float %tmp427, ptr addrspace(1) %tmp508, align 4594 %tmp509 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 61595 store float %tmp434, ptr addrspace(1) %tmp509, align 4596 %tmp510 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 62597 store float %tmp441, ptr addrspace(1) %tmp510, align 4598 %tmp511 = getelementptr inbounds float, ptr addrspace(1) %arg1, i64 63599 store float %tmp448, ptr addrspace(1) %tmp511, align 4600 ret void601}602 603; Function Attrs: nounwind readnone604declare float @llvm.fmuladd.f32(float, float, float) #0605 606attributes #0 = { nounwind readnone }607attributes #1 = { "amdgpu-wave-limiter"="true" }608