brintos

brintos / llvm-project-archived public Read only

0
0
Text · 6.4 KiB · 85cfd80 Raw
133 lines · plain
1# REQUIRES: asserts2# RUN: llc --mtriple=hexagon %s -run-pass=pipeliner -debug-only=pipeliner \3# RUN: -window-sched=force -filetype=null -verify-machineinstrs 2>&1 \4# RUN: | FileCheck %s5 6# CHECK: Best window offset is {{[0-9]+}} and Best II is {{[0-9]+}}.7# CHECK-LABEL: name: exp_approx_top_six8# CHECK: bb.5.loop_body:9# CHECK: dead %{{[0-9]*}}:hvxvr = V6_vaddw %{{[0-9]*}}, %{{[0-9]*}}10# CHECK: ENDLOOP011# CHECK: bb.6:12 13--- |14  define void @exp_approx_top_six(i32 %N, ptr noalias %x, ptr noalias %y) #0 {15  entry:16    %is_zero = icmp eq i32 %N, 017    br i1 %is_zero, label %exit, label %loop_header18 19  loop_header:20    %vec_one = tail call <32 x i32> @llvm.hexagon.V6.lvsplatw.128B(i32 1065353216)21    %vec_half = tail call <32 x i32> @llvm.hexagon.V6.lvsplatw.128B(i32 1056964608)22    %vec_sixth = tail call <32 x i32> @llvm.hexagon.V6.lvsplatw.128B(i32 1042983595)23    %vec_24th = tail call <32 x i32> @llvm.hexagon.V6.lvsplatw.128B(i32 1026206379)24    %vec_120th = tail call <32 x i32> @llvm.hexagon.V6.lvsplatw.128B(i32 1007192201)25    %vec_720th = tail call <32 x i32> @llvm.hexagon.V6.lvsplatw.128B(i32 985008993)26    br label %loop_body27 28  exit:29    ret void30 31  loop_body:32    %lsr.iv1 = phi ptr [ %cgep3, %loop_body ], [ %x, %loop_header ]33    %lsr.iv = phi ptr [ %cgep, %loop_body ], [ %y, %loop_header ]34    %index = phi i32 [ 0, %loop_header ], [ %index_next, %loop_body ]35    %vec_input = load <32 x i32>, ptr %lsr.iv1, align 12836    %vec_input_pow_2 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_input, <32 x i32> %vec_input)37    %vec_input_pow_3 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_input_pow_2, <32 x i32> %vec_input)38    %vec_input_pow_4 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_input_pow_3, <32 x i32> %vec_input)39    %vec_input_pow_5 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_input_pow_4, <32 x i32> %vec_input)40    %vec_input_pow_6 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_input_pow_5, <32 x i32> %vec_input)41    %vec_exp_approx_1 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_half, <32 x i32> %vec_input_pow_2)42    %vec_exp_approx_2 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_sixth, <32 x i32> %vec_input_pow_3)43    %vec_exp_approx_3 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_24th, <32 x i32> %vec_input_pow_4)44    %vec_exp_approx_4 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_120th, <32 x i32> %vec_input_pow_5)45    %vec_exp_approx_5 = tail call <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32> %vec_720th, <32 x i32> %vec_input_pow_6)46    %vec_exp_sum_1 = tail call <32 x i32> @llvm.hexagon.V6.vaddw.128B(<32 x i32> %vec_one, <32 x i32> %vec_input)47    %vec_exp_sum_2 = tail call <32 x i32> @llvm.hexagon.V6.vaddw.128B(<32 x i32> %vec_exp_sum_1, <32 x i32> %vec_exp_approx_1)48    %vec_exp_sum_3 = tail call <32 x i32> @llvm.hexagon.V6.vaddw.128B(<32 x i32> %vec_exp_sum_2, <32 x i32> %vec_exp_approx_2)49    %vec_exp_sum_4 = tail call <32 x i32> @llvm.hexagon.V6.vaddw.128B(<32 x i32> %vec_exp_sum_3, <32 x i32> %vec_exp_approx_3)50    %vec_exp_sum_5 = tail call <32 x i32> @llvm.hexagon.V6.vaddw.128B(<32 x i32> %vec_exp_sum_4, <32 x i32> %vec_exp_approx_4)51    %vec_exp_result = tail call <32 x i32> @llvm.hexagon.V6.vaddw.128B(<32 x i32> %vec_exp_sum_5, <32 x i32> %vec_exp_approx_5)52    store <32 x i32> %vec_exp_result, ptr %lsr.iv, align 12853    %index_next = add nuw i32 %index, 3254    %loop_cond = icmp ult i32 %index_next, %N55    %cgep = getelementptr i8, ptr %lsr.iv, i32 12856    %cgep3 = getelementptr i8, ptr %lsr.iv1, i32 12857    br i1 %loop_cond, label %loop_body, label %exit58  }59 60  declare <32 x i32> @llvm.hexagon.V6.vmpyowh.rnd.128B(<32 x i32>, <32 x i32>)61  declare <32 x i32> @llvm.hexagon.V6.lvsplatw.128B(i32)62  declare <32 x i32> @llvm.hexagon.V6.vaddw.128B(<32 x i32>, <32 x i32>)63 64  attributes #0 = { "target-features"="+hvx-length128b,+hvxv69,+v66,-long-calls" }65...66---67name:            exp_approx_top_six68tracksRegLiveness: true69body:             |70  bb.0.entry:71    successors: %bb.2(0x30000000), %bb.1(0x50000000)72    liveins: $r0, $r1, $r273 74    %0:intregs = COPY $r275    %1:intregs = COPY $r176    %2:intregs = COPY $r077    %3:predregs = C2_cmpeqi %2, 078    J2_jumpt killed %3, %bb.2, implicit-def dead $pc79    J2_jump %bb.1, implicit-def dead $pc80 81  bb.1.loop_header:82    successors: %bb.3(0x80000000)83 84    %4:intregs = A2_tfrsi 106535321685    %5:hvxvr = V6_lvsplatw killed %486    %6:intregs = A2_tfrsi 105696460887    %7:hvxvr = V6_lvsplatw killed %688    %8:intregs = A2_tfrsi 104298359589    %9:hvxvr = V6_lvsplatw killed %890    %10:intregs = A2_tfrsi 102620637991    %11:hvxvr = V6_lvsplatw killed %1092    %12:intregs = A2_tfrsi 100719220193    %13:hvxvr = V6_lvsplatw killed %1294    %14:intregs = A2_tfrsi 98500899395    %15:hvxvr = V6_lvsplatw killed %1496    %16:intregs = A2_addi %2, 3197    %17:intregs = S2_lsr_i_r %16, 598    %18:intregs = COPY %1799    J2_loop0r %bb.3, %18, implicit-def $lc0, implicit-def $sa0, implicit-def $usr100    J2_jump %bb.3, implicit-def dead $pc101 102  bb.2.exit:103    PS_jmpret $r31, implicit-def dead $pc104 105  bb.3.loop_body (machine-block-address-taken):106    successors: %bb.3(0x7c000000), %bb.2(0x04000000)107 108    %19:intregs = PHI %1, %bb.1, %20, %bb.3109    %21:intregs = PHI %0, %bb.1, %22, %bb.3110    %23:hvxvr, %20:intregs = V6_vL32b_pi %19, 128 :: (load (s1024) from %ir.lsr.iv1)111    %24:hvxvr = V6_vmpyowh_rnd %23, %23112    %25:hvxvr = V6_vmpyowh_rnd %24, %23113    %26:hvxvr = V6_vmpyowh_rnd %25, %23114    %27:hvxvr = V6_vmpyowh_rnd %26, %23115    %28:hvxvr = V6_vmpyowh_rnd %27, %23116    %29:hvxvr = V6_vmpyowh_rnd %7, %24117    %30:hvxvr = V6_vmpyowh_rnd %9, %25118    %31:hvxvr = V6_vmpyowh_rnd %11, %26119    %32:hvxvr = V6_vmpyowh_rnd %13, %27120    %33:hvxvr = V6_vmpyowh_rnd %15, killed %28121    %34:hvxvr = V6_vaddw %5, %23122    %35:hvxvr = V6_vaddw killed %34, killed %29123    %36:hvxvr = V6_vaddw killed %35, killed %30124    %37:hvxvr = V6_vaddw killed %36, killed %31125    %38:hvxvr = V6_vaddw killed %37, killed %32126    %39:hvxvr = V6_vaddw %38, %33127    ; To check the dead virtual register within loop kernel.128    dead %40:hvxvr = V6_vaddw killed %38, killed %33129    %22:intregs = V6_vS32b_pi %21, 128, killed %39 :: (store (s1024) into %ir.lsr.iv)130    ENDLOOP0 %bb.3, implicit-def $pc, implicit-def $lc0, implicit $sa0, implicit $lc0131    J2_jump %bb.2, implicit-def dead $pc132...133