Polly enabled optimizations in XLA for mat_add.ll

optimized_mat_add.ll

 ; ModuleID = '__compute_module'
 source_filename = "__compute_module"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux_gnu"
 
 ; Function Attrs: norecurse nounwind
 define void @"cluster_2[_XlaCompiledKernel=true,_XlaNumConstantArgs=0].v5"(i8* nocapture align 16 dereferenceable(3136) %retval, i8* noalias nocapture readnone %run_options, i8** noalias nocapture readonly %params, i8** noalias nocapture readnone %temps, i64* noalias nocapture readnone %prof_counters) local_unnamed_addr #0 {
 entry:
   %0 = bitcast i8** %params to [1 x [784 x float]]**
   %1 = load [1 x [784 x float]]*, [1 x [784 x float]]** %0, align 8, !tbaa !0, !dereferenceable !3, !align !4
   %2 = getelementptr inbounds i8*, i8** %params, i64 1
   %3 = bitcast i8** %2 to float**
   %4 = load float*, float** %3, align 8, !tbaa !5, !dereferenceable !7, !align !8
   %5 = bitcast i8* %retval to [1 x [784 x float]]*
   %6 = load float, float* %4, align 8, !tbaa !9, !invariant.load !11, !noalias !12
   %broadcast.splatinsert10 = insertelement <8 x float> undef, float %6, i32 0
   %broadcast.splat11 = shufflevector <8 x float> %broadcast.splatinsert10, <8 x float> undef, <8 x i32> zeroinitializer
   br label %vector.body
 
 vector.body:                                      ; preds = %vector.body, %entry
   %index = phi i64 [ 0, %entry ], [ %index.next.2, %vector.body ]
   %7 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 %index
   %8 = bitcast float* %7 to <8 x float>*
   %wide.load = load <8 x float>, <8 x float>* %8, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %9 = getelementptr float, float* %7, i64 8
   %10 = bitcast float* %9 to <8 x float>*
   %wide.load7 = load <8 x float>, <8 x float>* %10, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %11 = getelementptr float, float* %7, i64 16
   %12 = bitcast float* %11 to <8 x float>*
   %wide.load8 = load <8 x float>, <8 x float>* %12, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %13 = getelementptr float, float* %7, i64 24
   %14 = bitcast float* %13 to <8 x float>*
   %wide.load9 = load <8 x float>, <8 x float>* %14, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %15 = fadd fast <8 x float> %broadcast.splat11, %wide.load
   %16 = fadd fast <8 x float> %broadcast.splat11, %wide.load7
   %17 = fadd fast <8 x float> %broadcast.splat11, %wide.load8
   %18 = fadd fast <8 x float> %broadcast.splat11, %wide.load9
   %19 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %5, i64 0, i64 0, i64 %index
   %20 = bitcast float* %19 to <8 x float>*
   store <8 x float> %15, <8 x float>* %20, align 16, !tbaa !15, !alias.scope !12
   %21 = getelementptr float, float* %19, i64 8
   %22 = bitcast float* %21 to <8 x float>*
   store <8 x float> %16, <8 x float>* %22, align 16, !tbaa !15, !alias.scope !12
   %23 = getelementptr float, float* %19, i64 16
   %24 = bitcast float* %23 to <8 x float>*
   store <8 x float> %17, <8 x float>* %24, align 16, !tbaa !15, !alias.scope !12
   %25 = getelementptr float, float* %19, i64 24
   %26 = bitcast float* %25 to <8 x float>*
   store <8 x float> %18, <8 x float>* %26, align 16, !tbaa !15, !alias.scope !12
   %index.next = add nuw nsw i64 %index, 32
   %27 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 %index.next
   %28 = bitcast float* %27 to <8 x float>*
   %wide.load.1 = load <8 x float>, <8 x float>* %28, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %29 = getelementptr float, float* %27, i64 8
   %30 = bitcast float* %29 to <8 x float>*
   %wide.load7.1 = load <8 x float>, <8 x float>* %30, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %31 = getelementptr float, float* %27, i64 16
   %32 = bitcast float* %31 to <8 x float>*
   %wide.load8.1 = load <8 x float>, <8 x float>* %32, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %33 = getelementptr float, float* %27, i64 24
   %34 = bitcast float* %33 to <8 x float>*
   %wide.load9.1 = load <8 x float>, <8 x float>* %34, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %35 = fadd fast <8 x float> %broadcast.splat11, %wide.load.1
   %36 = fadd fast <8 x float> %broadcast.splat11, %wide.load7.1
   %37 = fadd fast <8 x float> %broadcast.splat11, %wide.load8.1
   %38 = fadd fast <8 x float> %broadcast.splat11, %wide.load9.1
   %39 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %5, i64 0, i64 0, i64 %index.next
   %40 = bitcast float* %39 to <8 x float>*
   store <8 x float> %35, <8 x float>* %40, align 16, !tbaa !15, !alias.scope !12
   %41 = getelementptr float, float* %39, i64 8
   %42 = bitcast float* %41 to <8 x float>*
   store <8 x float> %36, <8 x float>* %42, align 16, !tbaa !15, !alias.scope !12
   %43 = getelementptr float, float* %39, i64 16
   %44 = bitcast float* %43 to <8 x float>*
   store <8 x float> %37, <8 x float>* %44, align 16, !tbaa !15, !alias.scope !12
   %45 = getelementptr float, float* %39, i64 24
   %46 = bitcast float* %45 to <8 x float>*
   store <8 x float> %38, <8 x float>* %46, align 16, !tbaa !15, !alias.scope !12
   %index.next.1 = add nsw i64 %index, 64
   %47 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 %index.next.1
   %48 = bitcast float* %47 to <8 x float>*
   %wide.load.2 = load <8 x float>, <8 x float>* %48, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %49 = getelementptr float, float* %47, i64 8
   %50 = bitcast float* %49 to <8 x float>*
   %wide.load7.2 = load <8 x float>, <8 x float>* %50, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %51 = getelementptr float, float* %47, i64 16
   %52 = bitcast float* %51 to <8 x float>*
   %wide.load8.2 = load <8 x float>, <8 x float>* %52, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %53 = getelementptr float, float* %47, i64 24
   %54 = bitcast float* %53 to <8 x float>*
   %wide.load9.2 = load <8 x float>, <8 x float>* %54, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %55 = fadd fast <8 x float> %broadcast.splat11, %wide.load.2
   %56 = fadd fast <8 x float> %broadcast.splat11, %wide.load7.2
   %57 = fadd fast <8 x float> %broadcast.splat11, %wide.load8.2
   %58 = fadd fast <8 x float> %broadcast.splat11, %wide.load9.2
   %59 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %5, i64 0, i64 0, i64 %index.next.1
   %60 = bitcast float* %59 to <8 x float>*
   store <8 x float> %55, <8 x float>* %60, align 16, !tbaa !15, !alias.scope !12
   %61 = getelementptr float, float* %59, i64 8
   %62 = bitcast float* %61 to <8 x float>*
   store <8 x float> %56, <8 x float>* %62, align 16, !tbaa !15, !alias.scope !12
   %63 = getelementptr float, float* %59, i64 16
   %64 = bitcast float* %63 to <8 x float>*
   store <8 x float> %57, <8 x float>* %64, align 16, !tbaa !15, !alias.scope !12
   %65 = getelementptr float, float* %59, i64 24
   %66 = bitcast float* %65 to <8 x float>*
   store <8 x float> %58, <8 x float>* %66, align 16, !tbaa !15, !alias.scope !12
   %index.next.2 = add nsw i64 %index, 96
   %67 = icmp eq i64 %index.next.2, 768
   br i1 %67, label %loop_body.dim.1.preheader, label %vector.body, !llvm.loop !17
 
 loop_body.dim.1.preheader:                        ; preds = %vector.body
   br label %loop_body.dim.1
 
 loop_body.dim.1:                                  ; preds = %loop_body.dim.1.preheader
   %68 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 768
   %69 = load float, float* %68, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %70 = fadd fast float %6, %69
   %71 = getelementptr inbounds i8, i8* %retval, i64 3072
   %72 = bitcast i8* %71 to float*
   store float %70, float* %72, align 16, !tbaa !15, !alias.scope !12
   %73 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 769
   %74 = load float, float* %73, align 4, !tbaa !15, !invariant.load !11, !noalias !12
   %75 = fadd fast float %6, %74
   %76 = getelementptr inbounds i8, i8* %retval, i64 3076
   %77 = bitcast i8* %76 to float*
   store float %75, float* %77, align 4, !tbaa !15, !alias.scope !12
   %78 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 770
   %79 = load float, float* %78, align 8, !tbaa !15, !invariant.load !11, !noalias !12
   %80 = fadd fast float %6, %79
   %81 = getelementptr inbounds i8, i8* %retval, i64 3080
   %82 = bitcast i8* %81 to float*
   store float %80, float* %82, align 8, !tbaa !15, !alias.scope !12
   %83 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 771
   %84 = load float, float* %83, align 4, !tbaa !15, !invariant.load !11, !noalias !12
   %85 = fadd fast float %6, %84
   %86 = getelementptr inbounds i8, i8* %retval, i64 3084
   %87 = bitcast i8* %86 to float*
   store float %85, float* %87, align 4, !tbaa !15, !alias.scope !12
   %88 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 772
   %89 = load float, float* %88, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %90 = fadd fast float %6, %89
   %91 = getelementptr inbounds i8, i8* %retval, i64 3088
   %92 = bitcast i8* %91 to float*
   store float %90, float* %92, align 16, !tbaa !15, !alias.scope !12
   %93 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 773
   %94 = load float, float* %93, align 4, !tbaa !15, !invariant.load !11, !noalias !12
   %95 = fadd fast float %6, %94
   %96 = getelementptr inbounds i8, i8* %retval, i64 3092
   %97 = bitcast i8* %96 to float*
   store float %95, float* %97, align 4, !tbaa !15, !alias.scope !12
   %98 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 774
   %99 = load float, float* %98, align 8, !tbaa !15, !invariant.load !11, !noalias !12
   %100 = fadd fast float %6, %99
   %101 = getelementptr inbounds i8, i8* %retval, i64 3096
   %102 = bitcast i8* %101 to float*
   store float %100, float* %102, align 8, !tbaa !15, !alias.scope !12
   %103 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 775
   %104 = load float, float* %103, align 4, !tbaa !15, !invariant.load !11, !noalias !12
   %105 = fadd fast float %6, %104
   %106 = getelementptr inbounds i8, i8* %retval, i64 3100
   %107 = bitcast i8* %106 to float*
   store float %105, float* %107, align 4, !tbaa !15, !alias.scope !12
   %108 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 776
   %109 = load float, float* %108, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %110 = fadd fast float %6, %109
   %111 = getelementptr inbounds i8, i8* %retval, i64 3104
   %112 = bitcast i8* %111 to float*
   store float %110, float* %112, align 16, !tbaa !15, !alias.scope !12
   %113 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 777
   %114 = load float, float* %113, align 4, !tbaa !15, !invariant.load !11, !noalias !12
   %115 = fadd fast float %6, %114
   %116 = getelementptr inbounds i8, i8* %retval, i64 3108
   %117 = bitcast i8* %116 to float*
   store float %115, float* %117, align 4, !tbaa !15, !alias.scope !12
   %118 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 778
   %119 = load float, float* %118, align 8, !tbaa !15, !invariant.load !11, !noalias !12
   %120 = fadd fast float %6, %119
   %121 = getelementptr inbounds i8, i8* %retval, i64 3112
   %122 = bitcast i8* %121 to float*
   store float %120, float* %122, align 8, !tbaa !15, !alias.scope !12
   %123 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 779
   %124 = load float, float* %123, align 4, !tbaa !15, !invariant.load !11, !noalias !12
   %125 = fadd fast float %6, %124
   %126 = getelementptr inbounds i8, i8* %retval, i64 3116
   %127 = bitcast i8* %126 to float*
   store float %125, float* %127, align 4, !tbaa !15, !alias.scope !12
   %128 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 780
   %129 = load float, float* %128, align 16, !tbaa !15, !invariant.load !11, !noalias !12
   %130 = fadd fast float %6, %129
   %131 = getelementptr inbounds i8, i8* %retval, i64 3120
   %132 = bitcast i8* %131 to float*
   store float %130, float* %132, align 16, !tbaa !15, !alias.scope !12
   %133 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 781
   %134 = load float, float* %133, align 4, !tbaa !15, !invariant.load !11, !noalias !12
   %135 = fadd fast float %6, %134
   %136 = getelementptr inbounds i8, i8* %retval, i64 3124
   %137 = bitcast i8* %136 to float*
   store float %135, float* %137, align 4, !tbaa !15, !alias.scope !12
   %138 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 782
   %139 = load float, float* %138, align 8, !tbaa !15, !invariant.load !11, !noalias !12
   %140 = fadd fast float %6, %139
   %141 = getelementptr inbounds i8, i8* %retval, i64 3128
   %142 = bitcast i8* %141 to float*
   store float %140, float* %142, align 8, !tbaa !15, !alias.scope !12
   %143 = getelementptr inbounds [1 x [784 x float]], [1 x [784 x float]]* %1, i64 0, i64 0, i64 783
   %144 = load float, float* %143, align 4, !tbaa !15, !invariant.load !11, !noalias !12
   %145 = fadd fast float %6, %144
   %146 = getelementptr inbounds i8, i8* %retval, i64 3132
   %147 = bitcast i8* %146 to float*
   store float %145, float* %147, align 4, !tbaa !15, !alias.scope !12
   ret void
 }
 
 attributes #0 = { norecurse nounwind }
 
 !0 = !{!1, !1, i64 0}
 !1 = !{!"pointer-to element_type: F32 dimensions: 1 dimensions: 784 layout { minor_to_major: 1 minor_to_major: 0 }", !2}
 !2 = !{!"XLA TBAA"}
 !3 = !{i64 3136}
 !4 = !{i64 16}
 !5 = !{!6, !6, i64 0}
 !6 = !{!"pointer-to element_type: F32", !2}
 !7 = !{i64 4}
 !8 = !{i64 8}
 !9 = !{!10, !10, i64 0}
 !10 = !{!"element_type: F32", !2}
 !11 = !{}
 !12 = !{!13}
 !13 = !{!"buffer: 1", !14}
 !14 = distinct !{!14}
 !15 = !{!16, !16, i64 0}
 !16 = !{!"element_type: F32 dimensions: 1 dimensions: 784 layout { minor_to_major: 1 minor_to_major: 0 }", !2}
 !17 = distinct !{!17, !18, !19}
 !18 = !{!"llvm.loop.vectorize.width", i32 1}
 !19 = !{!"llvm.loop.interleave.count", i32 1}