111 lines · plain
1; REQUIRES: asserts2 3; RUN: llc -mtriple=amdgcn -mcpu=tonga -debug-only=machine-scheduler -filetype=null < %s 2>&1 | FileCheck %s4 5; We are only targeting one wave. Check that the machine scheduler doesn't use6; register pressure heuristics to prioritize any candidate instruction.7 8; CHECK-NOT: REG-CRIT9; CHECK-NOT: REG-EXCESS10 11define amdgpu_kernel void @load_fma_store(ptr addrspace(3) nocapture readonly %arg, ptr addrspace(3) nocapture %arg1) #1 {12bb:13 %tmp0 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 114 %tmp1 = load float, ptr addrspace(3) %tmp0, align 415 %tmp2 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 216 %tmp3 = load float, ptr addrspace(3) %tmp2, align 417 %tmp4 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 318 %tmp5 = load float, ptr addrspace(3) %tmp4, align 419 %tmp6 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 420 %tmp7 = load float, ptr addrspace(3) %tmp6, align 421 %tmp8 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 522 %tmp9 = load float, ptr addrspace(3) %tmp8, align 423 %tmp10 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 624 %tmp11 = load float, ptr addrspace(3) %tmp10, align 425 %tmp12 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 726 %tmp13 = load float, ptr addrspace(3) %tmp12, align 427 %tmp14 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 828 %tmp15 = load float, ptr addrspace(3) %tmp14, align 429 %tmp16 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 930 %tmp17 = load float, ptr addrspace(3) %tmp16, align 431 %tmp18 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1032 %tmp19 = load float, ptr addrspace(3) %tmp18, align 433 %tmp20 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1134 %tmp21 = load float, ptr addrspace(3) %tmp20, align 435 %tmp22 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1236 %tmp23 = load float, ptr addrspace(3) %tmp22, align 437 %tmp24 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1338 %tmp25 = load float, ptr addrspace(3) %tmp24, align 439 %tmp26 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1440 %tmp27 = load float, ptr addrspace(3) %tmp26, align 441 %tmp28 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1542 %tmp29 = load float, ptr addrspace(3) %tmp28, align 443 %tmp30 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1644 %tmp31 = load float, ptr addrspace(3) %tmp30, align 445 %tmp32 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1746 %tmp33 = load float, ptr addrspace(3) %tmp32, align 447 %tmp34 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1848 %tmp35 = load float, ptr addrspace(3) %tmp34, align 449 %tmp36 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 1950 %tmp37 = load float, ptr addrspace(3) %tmp36, align 451 %tmp38 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2052 %tmp39 = load float, ptr addrspace(3) %tmp38, align 453 %tmp40 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2154 %tmp41 = load float, ptr addrspace(3) %tmp40, align 455 %tmp42 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2256 %tmp43 = load float, ptr addrspace(3) %tmp42, align 457 %tmp44 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2358 %tmp45 = load float, ptr addrspace(3) %tmp44, align 459 %tmp46 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2460 %tmp47 = load float, ptr addrspace(3) %tmp46, align 461 %tmp48 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2562 %tmp49 = load float, ptr addrspace(3) %tmp48, align 463 %tmp50 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2664 %tmp51 = load float, ptr addrspace(3) %tmp50, align 465 %tmp52 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2766 %tmp53 = load float, ptr addrspace(3) %tmp52, align 467 %tmp54 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2868 %tmp55 = load float, ptr addrspace(3) %tmp54, align 469 %tmp56 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 2970 %tmp57 = load float, ptr addrspace(3) %tmp56, align 471 %tmp58 = getelementptr inbounds float, ptr addrspace(3) %arg, i32 3072 %tmp59 = load float, ptr addrspace(3) %tmp58, align 473 %tmp60 = tail call float @llvm.fmuladd.f32(float %tmp1, float %tmp3, float %tmp5)74 %tmp61 = tail call float @llvm.fmuladd.f32(float %tmp7, float %tmp9, float %tmp11)75 %tmp62 = tail call float @llvm.fmuladd.f32(float %tmp13, float %tmp15, float %tmp17)76 %tmp63 = tail call float @llvm.fmuladd.f32(float %tmp19, float %tmp21, float %tmp23)77 %tmp64 = tail call float @llvm.fmuladd.f32(float %tmp25, float %tmp27, float %tmp29)78 %tmp65 = tail call float @llvm.fmuladd.f32(float %tmp31, float %tmp33, float %tmp35)79 %tmp66 = tail call float @llvm.fmuladd.f32(float %tmp37, float %tmp39, float %tmp41)80 %tmp67 = tail call float @llvm.fmuladd.f32(float %tmp43, float %tmp45, float %tmp47)81 %tmp68 = tail call float @llvm.fmuladd.f32(float %tmp49, float %tmp51, float %tmp53)82 %tmp69 = tail call float @llvm.fmuladd.f32(float %tmp55, float %tmp57, float %tmp59)83 %tmp70 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 184 store float %tmp60, ptr addrspace(3) %tmp70, align 485 %tmp71 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 286 store float %tmp61, ptr addrspace(3) %tmp71, align 487 %tmp72 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 388 store float %tmp62, ptr addrspace(3) %tmp72, align 489 %tmp73 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 490 store float %tmp63, ptr addrspace(3) %tmp73, align 491 %tmp74 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 592 store float %tmp64, ptr addrspace(3) %tmp74, align 493 %tmp75 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 694 store float %tmp65, ptr addrspace(3) %tmp75, align 495 %tmp76 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 796 store float %tmp66, ptr addrspace(3) %tmp76, align 497 %tmp77 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 898 store float %tmp67, ptr addrspace(3) %tmp77, align 499 %tmp78 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 9100 store float %tmp68, ptr addrspace(3) %tmp78, align 4101 %tmp79 = getelementptr inbounds float, ptr addrspace(3) %arg1, i64 10102 store float %tmp69, ptr addrspace(3) %tmp79, align 4103 ret void104}105 106; Function Attrs: nounwind readnone107declare float @llvm.fmuladd.f32(float, float, float) #0108 109attributes #0 = { nounwind readnone }110attributes #1 = { "amdgpu-waves-per-eu"="1,1" "amdgpu-flat-work-group-size"="1,256" }111