ループ一回で 8*8 要素ずつこなしている感じぽい
SIMDの8とアンロール
define internal void @parallel_add(i8* nocapture align 16 dereferenceable(4194304) %retval, i8* noalias nocapture readnone %run_options, i8** noalias nocapture readonly %params, i8** noalias nocapture readnone %temps, i64* noalias nocapture readonly %dynamic_loop_bounds, i64* noalias nocapture readnone %prof_counters) #0 {
entry:
%0 = bitcast i8** %params to [1024 x [1024 x float]]**
%name.untyped2 = load [1024 x [1024 x float]]*, [1024 x [1024 x float]]** %0, align 8, !dereferenceable !0, !align !1
%1 = getelementptr inbounds i8*, i8** %params, i64 1
%2 = bitcast i8** %1 to [1024 x [1024 x float]]**
%name.1.untyped3 = load [1024 x [1024 x float]]*, [1024 x [1024 x float]]** %2, align 8, !dereferenceable !0, !align !1
%add.clone = bitcast i8* %retval to [1024 x [1024 x float]]*
%3 = load i64, i64* %dynamic_loop_bounds, align 8
%dynamic_loop_bound_1 = getelementptr i64, i64* %dynamic_loop_bounds, i64 1
%4 = load i64, i64* %dynamic_loop_bound_1, align 8
%5 = icmp ult i64 %3, %4
br i1 %5, label %vector.ph.preheader, label %add.clone.loop_exit.dim.0
vector.ph.preheader: ; preds = %entry
br label %vector.ph
vector.ph: ; preds = %vector.ph.preheader, %add.clone.loop_exit.dim.1
%add.clone.invar_address.dim.0.05 = phi i64 [ %invar.inc, %add.clone.loop_exit.dim.1 ], [ %3, %vector.ph.preheader ]
br label %vector.body
vector.body: ; preds = %vector.body, %vector.ph
%index = phi i64 [ 0, %vector.ph ], [ %index.next.1, %vector.body ]
%6 = getelementptr inbounds [1024 x [1024 x float]], [1024 x [1024 x float]]* %name.untyped2, i64 0, i64 %add.clone.invar_address.dim.0.05, i64 %index
%7 = bitcast float* %6 to <8 x float>*
%wide.load = load <8 x float>, <8 x float>* %7, align 16, !noalias !2
%8 = getelementptr float, float* %6, i64 8
%9 = bitcast float* %8 to <8 x float>*
%wide.load10 = load <8 x float>, <8 x float>* %9, align 16, !noalias !2
%10 = getelementptr float, float* %6, i64 16
%11 = bitcast float* %10 to <8 x float>*
%wide.load11 = load <8 x float>, <8 x float>* %11, align 16, !noalias !2
%12 = getelementptr float, float* %6, i64 24
%13 = bitcast float* %12 to <8 x float>*
%wide.load12 = load <8 x float>, <8 x float>* %13, align 16, !noalias !2
%14 = getelementptr inbounds [1024 x [1024 x float]], [1024 x [1024 x float]]* %name.1.untyped3, i64 0, i64 %add.clone.invar_address.dim.0.05, i64 %index
%15 = bitcast float* %14 to <8 x float>*
%wide.load13 = load <8 x float>, <8 x float>* %15, align 16, !noalias !2
%16 = getelementptr float, float* %14, i64 8
%17 = bitcast float* %16 to <8 x float>*
%wide.load14 = load <8 x float>, <8 x float>* %17, align 16, !noalias !2
%18 = getelementptr float, float* %14, i64 16
%19 = bitcast float* %18 to <8 x float>*
%wide.load15 = load <8 x float>, <8 x float>* %19, align 16, !noalias !2
%20 = getelementptr float, float* %14, i64 24
%21 = bitcast float* %20 to <8 x float>*
%wide.load16 = load <8 x float>, <8 x float>* %21, align 16, !noalias !2
%22 = fadd fast <8 x float> %wide.load13, %wide.load
%23 = fadd fast <8 x float> %wide.load14, %wide.load10
%24 = fadd fast <8 x float> %wide.load15, %wide.load11
%25 = fadd fast <8 x float> %wide.load16, %wide.load12
%26 = getelementptr inbounds [1024 x [1024 x float]], [1024 x [1024 x float]]* %add.clone, i64 0, i64 %add.clone.invar_address.dim.0.05, i64 %index
%27 = bitcast float* %26 to <8 x float>*
store <8 x float> %22, <8 x float>* %27, align 16, !alias.scope !2
%28 = getelementptr float, float* %26, i64 8
%29 = bitcast float* %28 to <8 x float>*
store <8 x float> %23, <8 x float>* %29, align 16, !alias.scope !2
%30 = getelementptr float, float* %26, i64 16
%31 = bitcast float* %30 to <8 x float>*
store <8 x float> %24, <8 x float>* %31, align 16, !alias.scope !2
%32 = getelementptr float, float* %26, i64 24
%33 = bitcast float* %32 to <8 x float>*
store <8 x float> %25, <8 x float>* %33, align 16, !alias.scope !2
%index.next = or i64 %index, 32
%34 = getelementptr inbounds [1024 x [1024 x float]], [1024 x [1024 x float]]* %name.untyped2, i64 0, i64 %add.clone.invar_address.dim.0.05, i64 %index.next
%35 = bitcast float* %34 to <8 x float>*
%wide.load.1 = load <8 x float>, <8 x float>* %35, align 16, !noalias !2
%36 = getelementptr float, float* %34, i64 8
%37 = bitcast float* %36 to <8 x float>*
%wide.load10.1 = load <8 x float>, <8 x float>* %37, align 16, !noalias !2
%38 = getelementptr float, float* %34, i64 16
%39 = bitcast float* %38 to <8 x float>*
%wide.load11.1 = load <8 x float>, <8 x float>* %39, align 16, !noalias !2
%40 = getelementptr float, float* %34, i64 24
%41 = bitcast float* %40 to <8 x float>*
%wide.load12.1 = load <8 x float>, <8 x float>* %41, align 16, !noalias !2
%42 = getelementptr inbounds [1024 x [1024 x float]], [1024 x [1024 x float]]* %name.1.untyped3, i64 0, i64 %add.clone.invar_address.dim.0.05, i64 %index.next
%43 = bitcast float* %42 to <8 x float>*
%wide.load13.1 = load <8 x float>, <8 x float>* %43, align 16, !noalias !2
%44 = getelementptr float, float* %42, i64 8
%45 = bitcast float* %44 to <8 x float>*
%wide.load14.1 = load <8 x float>, <8 x float>* %45, align 16, !noalias !2
%46 = getelementptr float, float* %42, i64 16
%47 = bitcast float* %46 to <8 x float>*
%wide.load15.1 = load <8 x float>, <8 x float>* %47, align 16, !noalias !2
%48 = getelementptr float, float* %42, i64 24
%49 = bitcast float* %48 to <8 x float>*
%wide.load16.1 = load <8 x float>, <8 x float>* %49, align 16, !noalias !2
%50 = fadd fast <8 x float> %wide.load13.1, %wide.load.1
%51 = fadd fast <8 x float> %wide.load14.1, %wide.load10.1
%52 = fadd fast <8 x float> %wide.load15.1, %wide.load11.1
%53 = fadd fast <8 x float> %wide.load16.1, %wide.load12.1
%54 = getelementptr inbounds [1024 x [1024 x float]], [1024 x [1024 x float]]* %add.clone, i64 0, i64 %add.clone.invar_address.dim.0.05, i64 %index.next
%55 = bitcast float* %54 to <8 x float>*
store <8 x float> %50, <8 x float>* %55, align 16, !alias.scope !2
%56 = getelementptr float, float* %54, i64 8
%57 = bitcast float* %56 to <8 x float>*
store <8 x float> %51, <8 x float>* %57, align 16, !alias.scope !2
%58 = getelementptr float, float* %54, i64 16
%59 = bitcast float* %58 to <8 x float>*
store <8 x float> %52, <8 x float>* %59, align 16, !alias.scope !2
%60 = getelementptr float, float* %54, i64 24
%61 = bitcast float* %60 to <8 x float>*
store <8 x float> %53, <8 x float>* %61, align 16, !alias.scope !2
%index.next.1 = add nuw nsw i64 %index, 64
%62 = icmp eq i64 %index.next.1, 1024
br i1 %62, label %add.clone.loop_exit.dim.1, label %vector.body, !llvm.loop !5