conv_optimized.ll

; ModuleID = 'cluster_2.ll'
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"

@0 = private constant [1 x [28 x [28 x [1 x float]]]] [[28 x [28 x [1 x float]]] [[28 x [1 x float]] zeroinitializer, [28 x [1 x float]] zeroinitializer, [28 x [1 x float]] zeroinitializer, [28 x [1 x float]] zeroinitializer, [28 x [1 x float]] zeroinitializer, [28 x [1 x float]] zeroinitializer, [28 x [1 x float]] zeroinitializer, [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FD8585880000000], [1 x float] [float 0x3FD8181840000000], [1 x float] [float 0x3FD3535360000000], [1 x float] [float 0x3FDD9D9DC0000000], [1 x float] [float 0x3FCE9E9EC0000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FD69696A0000000], [1 x float] [float 0x3FE1515160000000], [1 x float] [float 0x3FED7D7DA0000000], [1 x float] [float 0x3FED7D7DA0000000], [1 x float] [float 0x3FED7D7DA0000000], [1 x float] [float 0x3FED7D7DA0000000], [1 x float] [float 0x3FED7D7DA0000000], [1 x float] [float 0x3FED7D7DA0000000], [1 x float] [float 0x3FEF7F7FA0000000], [1 x float] [float 0x3FEF7F7FA0000000], [1 x float] [float 0x3FEF1F1F40000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEEBEBEE0000000], [1 x float] [float 0x3FED7D7DA0000000], [1 x float] [float 0x3FE7D7D7E0000000], [1 x float] [float 0x3FB5151520000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FE19191A0000000], [1 x float] [float 0x3FEF7F7FA0000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FE7B7B7C0000000], [1 x float] [float 0x3FB7171720000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FEC5C5C80000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEA1A1A40000000], [1 x float] [float 0x3FE8F8F920000000], [1 x float] [float 0x3FE8F8F920000000], [1 x float] [float 0x3FE8F8F920000000], [1 x float] [float 0x3FE8F8F920000000], [1 x float] [float 0x3FE1717180000000], [1 x float] [float 0x3FCE9E9EC0000000], [1 x float] [float 0x3FCE9E9EC0000000], [1 x float] [float 0x3FCE9E9EC0000000], [1 x float] [float 0x3FCE9E9EC0000000], [1 x float] [float 0x3FCE9E9EC0000000], [1 x float] [float 0x3FE0101020000000], [1 x float] [float 0x3FEBDBDC00000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FE7B7B7C0000000], [1 x float] [float 0x3FB5151520000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FC3131320000000], [1 x float] [float 0x3FD49494A0000000], [1 x float] [float 0x3FAA1A1A40000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FC1111120000000], [1 x float] [float 0x3FEABABAE0000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FDCDCDD00000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FD5151520000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FED5D5D80000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FD5151520000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FED5D5D80000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FDA9A9AC0000000], [1 x float] [float 0x3FE3B3B3C0000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEE7E7EA0000000], [1 x float] [float 0x3FC99999C0000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FB9191940000000], [1 x float] [float 0x3FDD5D5D80000000], [1 x float] [float 0x3FEC9C9CC0000000], [1 x float] [float 0x3FEC9C9CC0000000], [1 x float] [float 0x3FEC9C9CC0000000], [1 x float] [float 0x3FEFBFBFE0000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEE1E1E40000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FD1111120000000], [1 x float] [float 0x3FDDDDDE00000000], [1 x float] [float 0x3FEB9B9BC0000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FE1D1D1E0000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FC29292A0000000], [1 x float] [float 0x3FE7777780000000], [1 x float] [float 0x3FEFBFBFE0000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEBFBFC20000000], [1 x float] [float 0x3FE9D9DA00000000], [1 x float] [float 0x3FE9D9DA00000000], [1 x float] [float 0x3FD2D2D2E0000000], [1 x float] [float 0x3FD1111120000000], [1 x float] [float 0x3FEAFAFB20000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FDD5D5D80000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FDC5C5C80000000], [1 x float] [float 0x3FEB7B7BA0000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEE5E5E80000000], [1 x float] [float 0x3FEC7C7CA0000000], [1 x float] [float 0x3FDCDCDD00000000], [1 x float] [float 0x3FD6565660000000], [1 x float] [float 0x3FBF1F1F40000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FE9191940000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEE3E3E60000000], [1 x float] [float 0x3FC49494A0000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FE5353540000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FE6161620000000], [1 x float] [float 0x3FCF1F1F40000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FC8181840000000], [1 x float] [float 0x3FECFCFD20000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FED5D5D80000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FB2121220000000], [1 x float] [float 0x3FDF1F1F40000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FD5151520000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FE4D4D4E0000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FE1717180000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEDDDDE00000000], [1 x float] [float 0x3FCC9C9CC0000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FEA5A5A80000000], [1 x float] [float 0x3FEF5F5F80000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FE5151520000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FEE5E5E80000000], [1 x float] [float 0x3FEFDFE000000000], [1 x float] [float 0x3FEDFDFE20000000], [1 x float] [float 0x3FCC9C9CC0000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3FD6565660000000], [1 x float] [float 0x3FEF7F7FA0000000], [1 x float] [float 0x3FEE3E3E60000000], [1 x float] [float 0x3FD59595A0000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3F94141420000000], [1 x float] [float 0x3FE9D9DA00000000], [1 x float] [float 0x3FEEDEDF00000000], [1 x float] [float 0x3FE3B3B3C0000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] [[1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] [float 0x3F90101020000000], [1 x float] [float 0x3FDD5D5D80000000], [1 x float] [float 0x3FD1515160000000], [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer, [1 x float] zeroinitializer], [28 x [1 x float]] zeroinitializer]]

define void @"cluster_2[_XlaCompiledKernel=true,_XlaNumConstantArgs=0].v7"(i8* align 16 dereferenceable(100352) %retval, i8* noalias %run_options, i8** noalias %params, i8** noalias %temps, i64* noalias %prof_counters) #0 {
entry:
  %convolution_sum_address.3.phiops = alloca float
  %.s2a51 = alloca i64
  %.s2a = alloca i64
  %convolution_sum_address.2.s2a = alloca float
  %convolution_sum_address.2.phiops = alloca float
  %convolution_sum_address.1.s2a = alloca float
  %convolution_sum_address.1.phiops = alloca float
  %convolution_sum_address.0.s2a = alloca float
  %convolution_sum_address.0.phiops = alloca float
  %.preload.s2a2 = alloca i8*
  %.preload.s2a = alloca i8*
  %0 = getelementptr inbounds i8*, i8** %params, i64 1
  %1 = load i8*, i8** %0, !tbaa !0, !dereferenceable !3, !align !4
  %2 = bitcast i8* %1 to [5 x [5 x [1 x [32 x float]]]]*
  %3 = getelementptr inbounds i8*, i8** %temps, i64 0
  %4 = load i8*, i8** %3, !tbaa !5, !invariant.load !7, !dereferenceable !8, !align !4
  %5 = bitcast i8* %4 to [1 x [28 x [28 x [32 x float]]]]*
  br label %polly.split_new_and_old

polly.split_new_and_old:                          ; preds = %entry
  br label %polly.preload.begin

polly.preload.begin:                              ; preds = %polly.split_new_and_old
  %polly.access.temps = getelementptr i8*, i8** %temps, i64 1
  %polly.access.temps.load = load i8*, i8** %polly.access.temps
  store i8* %polly.access.temps.load, i8** %.preload.s2a
  %polly.access.params = getelementptr i8*, i8** %params, i64 0
  %polly.access.params.load = load i8*, i8** %polly.access.params
  store i8* %polly.access.params.load, i8** %.preload.s2a2
  br i1 true, label %polly.start, label %loop_header.dim.0.pre_entry_bb

loop_header.dim.0.pre_entry_bb:                   ; preds = %polly.preload.begin
  br label %loop_header.dim.0

loop_header.dim.0:                                ; preds = %loop_header.dim.0.pre_entry_bb, %loop_exit.dim.1
  %invar_address.dim.0.0 = phi i64 [ %invar.inc, %loop_exit.dim.1 ], [ 0, %loop_header.dim.0.pre_entry_bb ]
  %6 = icmp uge i64 %invar_address.dim.0.0, 1
  br i1 %6, label %loop_exit.dim.0, label %loop_body.dim.0

loop_body.dim.0:                                  ; preds = %loop_header.dim.0
  br label %loop_header.dim.1

loop_header.dim.1:                                ; preds = %loop_exit.dim.2, %loop_body.dim.0
  %invar_address.dim.1.0 = phi i64 [ 0, %loop_body.dim.0 ], [ %invar.inc1, %loop_exit.dim.2 ]
  %7 = icmp uge i64 %invar_address.dim.1.0, 28
  br i1 %7, label %loop_exit.dim.1, label %loop_body.dim.1

loop_body.dim.1:                                  ; preds = %loop_header.dim.1
  br label %loop_header.dim.2

loop_header.dim.2:                                ; preds = %loop_exit.dim.3, %loop_body.dim.1
  %invar_address.dim.2.0 = phi i64 [ 0, %loop_body.dim.1 ], [ %invar.inc2, %loop_exit.dim.3 ]
  %8 = icmp uge i64 %invar_address.dim.2.0, 28
  br i1 %8, label %loop_exit.dim.2, label %loop_body.dim.2

loop_body.dim.2:                                  ; preds = %loop_header.dim.2
  br label %loop_header.dim.3

loop_header.dim.3:                                ; preds = %loop_exit.k0, %loop_body.dim.2
  %invar_address.dim.3.0 = phi i64 [ 0, %loop_body.dim.2 ], [ %invar.inc3, %loop_exit.k0 ]
  %9 = icmp uge i64 %invar_address.dim.3.0, 32
  br i1 %9, label %loop_exit.dim.3, label %loop_body.dim.3

loop_body.dim.3:                                  ; preds = %loop_header.dim.3
  br label %loop_header.k0

loop_header.k0:                                   ; preds = %loop_exit.k1, %loop_body.dim.3
  %invar_address.k0.0 = phi i64 [ 0, %loop_body.dim.3 ], [ %invar.inc4, %loop_exit.k1 ]
  %convolution_sum_address.0 = phi float [ 0.000000e+00, %loop_body.dim.3 ], [ %convolution_sum_address.1, %loop_exit.k1 ]
  %10 = icmp uge i64 %invar_address.k0.0, 5
  br i1 %10, label %loop_exit.k0, label %loop_body.k0

loop_body.k0:                                     ; preds = %loop_header.k0
  br label %loop_header.k1

loop_header.k1:                                   ; preds = %loop_exit.iz, %loop_body.k0
  %invar_address.k1.0 = phi i64 [ 0, %loop_body.k0 ], [ %invar.inc5, %loop_exit.iz ]
  %convolution_sum_address.1 = phi float [ %convolution_sum_address.0, %loop_body.k0 ], [ %convolution_sum_address.2, %loop_exit.iz ]
  %11 = icmp uge i64 %invar_address.k1.0, 5
  br i1 %11, label %loop_exit.k1, label %loop_body.k1

loop_body.k1:                                     ; preds = %loop_header.k1
  br label %loop_header.iz

loop_header.iz:                                   ; preds = %in-bounds-after, %loop_body.k1
  %invar_address.iz.0 = phi i64 [ 0, %loop_body.k1 ], [ %invar.inc6, %in-bounds-after ]
  %convolution_sum_address.2 = phi float [ %convolution_sum_address.1, %loop_body.k1 ], [ %convolution_sum_address.3, %in-bounds-after ]
  %12 = icmp uge i64 %invar_address.iz.0, 1
  br i1 %12, label %loop_exit.iz, label %loop_body.iz

loop_body.iz:                                     ; preds = %loop_header.iz
  %13 = mul nsw i64 %invar_address.dim.1.0, 1
  %14 = mul nsw i64 %invar_address.k0.0, 1
  %15 = add nsw i64 %13, %14
  %16 = sub nsw i64 %15, 2
  %17 = mul nsw i64 %invar_address.dim.2.0, 1
  %18 = mul nsw i64 %invar_address.k1.0, 1
  %19 = add nsw i64 %17, %18
  %20 = sub nsw i64 %19, 2
  %21 = icmp ult i64 %16, 28
  %22 = srem i64 %16, 1
  %23 = icmp eq i64 %22, 0
  %24 = and i1 %21, %23
  %25 = icmp ult i64 %20, 28
  %26 = srem i64 %20, 1
  %27 = icmp eq i64 %26, 0
  %28 = and i1 %25, %27
  %29 = and i1 %24, %28
  %30 = sdiv i64 %16, 1
  %31 = sdiv i64 %20, 1
  br i1 %29, label %in-bounds-true, label %in-bounds-false

in-bounds-after:                                  ; preds = %in-bounds-false, %in-bounds-true
  %convolution_sum_address.3 = phi float [ %44, %in-bounds-true ], [ %convolution_sum_address.2, %in-bounds-false ]
  %invar.inc6 = add nuw nsw i64 %invar_address.iz.0, 1
  br label %loop_header.iz

loop_exit.iz:                                     ; preds = %loop_header.iz
  %invar.inc5 = add nuw nsw i64 %invar_address.k1.0, 1
  br label %loop_header.k1

loop_exit.k1:                                     ; preds = %loop_header.k1
  %invar.inc4 = add nuw nsw i64 %invar_address.k0.0, 1
  br label %loop_header.k0

loop_exit.k0:                                     ; preds = %loop_header.k0
  %32 = getelementptr inbounds [1 x [28 x [28 x [32 x float]]]], [1 x [28 x [28 x [32 x float]]]]* %5, i64 0, i64 0, i64 %invar_address.dim.1.0, i64 %invar_address.dim.2.0, i64 %invar_address.dim.3.0
  store float %convolution_sum_address.0, float* %32, !tbaa !9, !alias.scope !11, !noalias !14
  %invar.inc3 = add nuw nsw i64 %invar_address.dim.3.0, 1
  br label %loop_header.dim.3

loop_exit.dim.3:                                  ; preds = %loop_header.dim.3
  %invar.inc2 = add nuw nsw i64 %invar_address.dim.2.0, 1
  br label %loop_header.dim.2

loop_exit.dim.2:                                  ; preds = %loop_header.dim.2
  %invar.inc1 = add nuw nsw i64 %invar_address.dim.1.0, 1
  br label %loop_header.dim.1

loop_exit.dim.1:                                  ; preds = %loop_header.dim.1
  %invar.inc = add nuw nsw i64 %invar_address.dim.0.0, 1
  br label %loop_header.dim.0

loop_exit.dim.0:                                  ; preds = %loop_header.dim.0
  %33 = getelementptr inbounds i8*, i8** %params, i64 0
  %34 = load i8*, i8** %33, !tbaa !16, !dereferenceable !18, !align !4
  %35 = bitcast i8* %34 to [32 x float]*
  %36 = getelementptr inbounds i8*, i8** %temps, i64 1
  %37 = load i8*, i8** %36, !tbaa !5, !invariant.load !7, !dereferenceable !8, !align !4
  %38 = bitcast i8* %37 to [1 x [28 x [28 x [32 x float]]]]*
  br label %loop_header.dim.08

in-bounds-true:                                   ; preds = %loop_body.iz
  %39 = getelementptr inbounds [5 x [5 x [1 x [32 x float]]]], [5 x [5 x [1 x [32 x float]]]]* %2, i64 0, i64 %invar_address.k0.0, i64 %invar_address.k1.0, i64 0, i64 %invar_address.dim.3.0
  %40 = load float, float* %39, !tbaa !19, !invariant.load !7, !noalias !11
  %41 = getelementptr inbounds [1 x [28 x [28 x [1 x float]]]], [1 x [28 x [28 x [1 x float]]]]* @0, i64 0, i64 0, i64 %30, i64 %31, i64 0
  %42 = load float, float* %41, !tbaa !21
  %43 = fmul fast float %42, %40
  %44 = fadd fast float %convolution_sum_address.2, %43
  br label %in-bounds-after

in-bounds-false:                                  ; preds = %loop_body.iz
  br label %in-bounds-after

loop_header.dim.08:                               ; preds = %loop_exit.dim.113, %loop_exit.dim.0
  %invar_address.dim.010.0 = phi i64 [ 0, %loop_exit.dim.0 ], [ %invar.inc12, %loop_exit.dim.113 ]
  %45 = icmp uge i64 %invar_address.dim.010.0, 1
  br i1 %45, label %loop_exit.dim.07, label %loop_body.dim.09

loop_body.dim.09:                                 ; preds = %loop_header.dim.08
  br label %loop_header.dim.114

loop_header.dim.114:                              ; preds = %loop_exit.dim.219, %loop_body.dim.09
  %invar_address.dim.116.0 = phi i64 [ 0, %loop_body.dim.09 ], [ %invar.inc18, %loop_exit.dim.219 ]
  %46 = icmp uge i64 %invar_address.dim.116.0, 28
  br i1 %46, label %loop_exit.dim.113, label %loop_body.dim.115

loop_body.dim.115:                                ; preds = %loop_header.dim.114
  br label %loop_header.dim.220

loop_header.dim.220:                              ; preds = %loop_exit.dim.325, %loop_body.dim.115
  %invar_address.dim.222.0 = phi i64 [ 0, %loop_body.dim.115 ], [ %invar.inc24, %loop_exit.dim.325 ]
  %47 = icmp uge i64 %invar_address.dim.222.0, 28
  br i1 %47, label %loop_exit.dim.219, label %loop_body.dim.221

loop_body.dim.221:                                ; preds = %loop_header.dim.220
  br label %loop_header.dim.326

loop_header.dim.326:                              ; preds = %loop_body.dim.327, %loop_body.dim.221
  %invar_address.dim.328.0 = phi i64 [ 0, %loop_body.dim.221 ], [ %invar.inc30, %loop_body.dim.327 ]
  %48 = icmp uge i64 %invar_address.dim.328.0, 32
  br i1 %48, label %loop_exit.dim.325, label %loop_body.dim.327

loop_body.dim.327:                                ; preds = %loop_header.dim.326
  %49 = getelementptr inbounds [32 x float], [32 x float]* %35, i64 0, i64 %invar_address.dim.328.0
  %50 = load float, float* %49, !tbaa !23, !invariant.load !7, !noalias !11
  %51 = getelementptr inbounds [1 x [28 x [28 x [32 x float]]]], [1 x [28 x [28 x [32 x float]]]]* %38, i64 0, i64 0, i64 %invar_address.dim.116.0, i64 %invar_address.dim.222.0, i64 %invar_address.dim.328.0
  store float %50, float* %51, !tbaa !9, !alias.scope !14, !noalias !11
  %invar.inc30 = add nuw nsw i64 %invar_address.dim.328.0, 1
  br label %loop_header.dim.326

loop_exit.dim.325:                                ; preds = %loop_header.dim.326
  %invar.inc24 = add nuw nsw i64 %invar_address.dim.222.0, 1
  br label %loop_header.dim.220

loop_exit.dim.219:                                ; preds = %loop_header.dim.220
  %invar.inc18 = add nuw nsw i64 %invar_address.dim.116.0, 1
  br label %loop_header.dim.114

loop_exit.dim.113:                                ; preds = %loop_header.dim.114
  %invar.inc12 = add nuw nsw i64 %invar_address.dim.010.0, 1
  br label %loop_header.dim.08

loop_exit.dim.07:                                 ; preds = %loop_header.dim.08
  %52 = bitcast i8* %retval to [1 x [28 x [28 x [32 x float]]]]*
  br label %loop_header.dim.032

loop_header.dim.032:                              ; preds = %loop_exit.dim.137, %loop_exit.dim.07
  %invar_address.dim.034.0 = phi i64 [ 0, %loop_exit.dim.07 ], [ %invar.inc36, %loop_exit.dim.137 ]
  %53 = icmp uge i64 %invar_address.dim.034.0, 1
  br i1 %53, label %polly.merge_new_and_old, label %loop_body.dim.033

loop_body.dim.033:                                ; preds = %loop_header.dim.032
  br label %loop_header.dim.138

loop_header.dim.138:                              ; preds = %loop_exit.dim.243, %loop_body.dim.033
  %invar_address.dim.140.0 = phi i64 [ 0, %loop_body.dim.033 ], [ %invar.inc42, %loop_exit.dim.243 ]
  %54 = icmp uge i64 %invar_address.dim.140.0, 28
  br i1 %54, label %loop_exit.dim.137, label %loop_body.dim.139

loop_body.dim.139:                                ; preds = %loop_header.dim.138
  br label %loop_header.dim.244

loop_header.dim.244:                              ; preds = %loop_exit.dim.349, %loop_body.dim.139
  %invar_address.dim.246.0 = phi i64 [ 0, %loop_body.dim.139 ], [ %invar.inc48, %loop_exit.dim.349 ]
  %55 = icmp uge i64 %invar_address.dim.246.0, 28
  br i1 %55, label %loop_exit.dim.243, label %loop_body.dim.245

loop_body.dim.245:                                ; preds = %loop_header.dim.244
  br label %loop_header.dim.350

loop_header.dim.350:                              ; preds = %loop_body.dim.351, %loop_body.dim.245
  %invar_address.dim.352.0 = phi i64 [ 0, %loop_body.dim.245 ], [ %invar.inc54, %loop_body.dim.351 ]
  %56 = icmp uge i64 %invar_address.dim.352.0, 32
  br i1 %56, label %loop_exit.dim.349, label %loop_body.dim.351

loop_body.dim.351:                                ; preds = %loop_header.dim.350
  %57 = getelementptr inbounds [1 x [28 x [28 x [32 x float]]]], [1 x [28 x [28 x [32 x float]]]]* %5, i64 0, i64 0, i64 %invar_address.dim.140.0, i64 %invar_address.dim.246.0, i64 %invar_address.dim.352.0
  %58 = load float, float* %57, !tbaa !9, !alias.scope !11, !noalias !14
  %59 = getelementptr inbounds [1 x [28 x [28 x [32 x float]]]], [1 x [28 x [28 x [32 x float]]]]* %38, i64 0, i64 0, i64 %invar_address.dim.140.0, i64 %invar_address.dim.246.0, i64 %invar_address.dim.352.0
  %60 = load float, float* %59, !tbaa !9, !alias.scope !14, !noalias !11
  %61 = fadd fast float %58, %60
  %62 = getelementptr inbounds [1 x [28 x [28 x [32 x float]]]], [1 x [28 x [28 x [32 x float]]]]* %52, i64 0, i64 0, i64 %invar_address.dim.140.0, i64 %invar_address.dim.246.0, i64 %invar_address.dim.352.0
  store float %61, float* %62, !tbaa !9, !alias.scope !11, !noalias !14
  %invar.inc54 = add nuw nsw i64 %invar_address.dim.352.0, 1
  br label %loop_header.dim.350

loop_exit.dim.349:                                ; preds = %loop_header.dim.350
  %invar.inc48 = add nuw nsw i64 %invar_address.dim.246.0, 1
  br label %loop_header.dim.244

loop_exit.dim.243:                                ; preds = %loop_header.dim.244
  %invar.inc42 = add nuw nsw i64 %invar_address.dim.140.0, 1
  br label %loop_header.dim.138

loop_exit.dim.137:                                ; preds = %loop_header.dim.138
  %invar.inc36 = add nuw nsw i64 %invar_address.dim.034.0, 1
  br label %loop_header.dim.032

polly.merge_new_and_old:                          ; preds = %polly.exiting, %loop_header.dim.032
  br label %loop_exit.dim.031

loop_exit.dim.031:                                ; preds = %polly.merge_new_and_old
  %prof_counter_computation = getelementptr i64, i64* %prof_counters, i64 0
  ret void

polly.start:                                      ; preds = %polly.preload.begin
  br label %polly.loop_preheader

polly.loop_exit:                                  ; preds = %polly.loop_exit5
  br label %polly.loop_preheader71

polly.loop_exit72:                                ; preds = %polly.loop_exit78
  br label %polly.loop_preheader96

polly.loop_exit97:                                ; preds = %polly.loop_exit103
  br label %polly.exiting

polly.exiting:                                    ; preds = %polly.loop_exit97
  br label %polly.merge_new_and_old

polly.loop_header:                                ; preds = %polly.loop_exit5, %polly.loop_preheader
  %polly.indvar = phi i64 [ 0, %polly.loop_preheader ], [ %polly.indvar_next, %polly.loop_exit5 ]
  br label %polly.loop_preheader4

polly.loop_exit5:                                 ; preds = %polly.loop_exit11
  %polly.indvar_next = add nsw i64 %polly.indvar, 1
  %polly.loop_cond = icmp sle i64 %polly.indvar_next, 27
  br i1 %polly.loop_cond, label %polly.loop_header, label %polly.loop_exit

polly.loop_preheader:                             ; preds = %polly.start
  br label %polly.loop_header

polly.loop_header3:                               ; preds = %polly.loop_exit11, %polly.loop_preheader4
  %polly.indvar6 = phi i64 [ 0, %polly.loop_preheader4 ], [ %polly.indvar_next7, %polly.loop_exit11 ]
  br label %polly.loop_preheader10

polly.loop_exit11:                                ; preds = %polly.stmt.loop_exit.k0
  %polly.indvar_next7 = add nsw i64 %polly.indvar6, 1
  %polly.loop_cond8 = icmp sle i64 %polly.indvar_next7, 27
  br i1 %polly.loop_cond8, label %polly.loop_header3, label %polly.loop_exit5

polly.loop_preheader4:                            ; preds = %polly.loop_header
  %63 = add i64 %polly.indvar, -2
  %64 = mul i64 %polly.indvar, 3584
  %scevgep66 = getelementptr i8, i8* %4, i64 %64
  br label %polly.loop_header3

polly.loop_header9:                               ; preds = %polly.stmt.loop_exit.k0, %polly.loop_preheader10
  %polly.indvar12 = phi i64 [ 0, %polly.loop_preheader10 ], [ %polly.indvar_next13, %polly.stmt.loop_exit.k0 ]
  br label %polly.stmt.loop_body.dim.3

polly.stmt.loop_body.dim.3:                       ; preds = %polly.loop_header9
  store float 0.000000e+00, float* %convolution_sum_address.0.phiops
  br label %polly.loop_preheader16

polly.loop_exit17:                                ; preds = %polly.merge
  br label %polly.stmt.loop_exit.k0

polly.stmt.loop_exit.k0:                          ; preds = %polly.loop_exit17
  %convolution_sum_address.0.s2a.reload65 = load float, float* %convolution_sum_address.0.s2a
  %65 = shl i64 %polly.indvar12, 2
  %scevgep68 = getelementptr i8, i8* %scevgep67, i64 %65
  %scevgep6869 = bitcast i8* %scevgep68 to float*
  store float %convolution_sum_address.0.s2a.reload65, float* %scevgep6869
  %polly.indvar_next13 = add nsw i64 %polly.indvar12, 1
  %polly.loop_cond14 = icmp sle i64 %polly.indvar_next13, 31
  br i1 %polly.loop_cond14, label %polly.loop_header9, label %polly.loop_exit11

polly.loop_preheader10:                           ; preds = %polly.loop_header3
  %66 = add i64 %polly.indvar6, -2
  %67 = shl i64 %polly.indvar6, 7
  %scevgep67 = getelementptr i8, i8* %scevgep66, i64 %67
  br label %polly.loop_header9

polly.loop_header15:                              ; preds = %polly.merge, %polly.loop_preheader16
  %polly.indvar18 = phi i64 [ 0, %polly.loop_preheader16 ], [ %polly.indvar_next19, %polly.merge ]
  br label %polly.stmt.loop_header.k0

polly.stmt.loop_header.k0:                        ; preds = %polly.loop_header15
  %convolution_sum_address.0.phiops.reload = load float, float* %convolution_sum_address.0.phiops
  store float %convolution_sum_address.0.phiops.reload, float* %convolution_sum_address.0.s2a
  br label %polly.cond

polly.cond:                                       ; preds = %polly.stmt.loop_header.k0
  %68 = icmp sle i64 %polly.indvar18, 4
  br i1 %68, label %polly.then, label %polly.else

polly.merge:                                      ; preds = %polly.else, %polly.stmt.loop_exit.k1
  %polly.indvar_next19 = add nsw i64 %polly.indvar18, 1
  %polly.loop_cond20 = icmp sle i64 %polly.indvar_next19, 5
  br i1 %polly.loop_cond20, label %polly.loop_header15, label %polly.loop_exit17

polly.loop_preheader16:                           ; preds = %polly.stmt.loop_body.dim.3
  %69 = shl i64 %polly.indvar12, 2
  %scevgep = getelementptr i8, i8* %1, i64 %69
  br label %polly.loop_header15

polly.then:                                       ; preds = %polly.cond
  br label %polly.stmt.loop_body.k0

polly.stmt.loop_body.k0:                          ; preds = %polly.then
  %convolution_sum_address.0.s2a.reload = load float, float* %convolution_sum_address.0.s2a
  store float %convolution_sum_address.0.s2a.reload, float* %convolution_sum_address.1.phiops
  br label %polly.loop_preheader22

polly.loop_exit23:                                ; preds = %polly.merge28
  br label %polly.stmt.loop_exit.k1

polly.stmt.loop_exit.k1:                          ; preds = %polly.loop_exit23
  %convolution_sum_address.1.s2a.reload64 = load float, float* %convolution_sum_address.1.s2a
  store float %convolution_sum_address.1.s2a.reload64, float* %convolution_sum_address.0.phiops
  br label %polly.merge

polly.else:                                       ; preds = %polly.cond
  br label %polly.merge

polly.loop_header21:                              ; preds = %polly.merge28, %polly.loop_preheader22
  %polly.indvar24 = phi i64 [ 0, %polly.loop_preheader22 ], [ %polly.indvar_next25, %polly.merge28 ]
  br label %polly.stmt.loop_header.k1

polly.stmt.loop_header.k1:                        ; preds = %polly.loop_header21
  %convolution_sum_address.1.phiops.reload = load float, float* %convolution_sum_address.1.phiops
  store float %convolution_sum_address.1.phiops.reload, float* %convolution_sum_address.1.s2a
  br label %polly.cond27

polly.cond27:                                     ; preds = %polly.stmt.loop_header.k1
  %70 = icmp sle i64 %polly.indvar24, 4
  br i1 %70, label %polly.then29, label %polly.else30

polly.merge28:                                    ; preds = %polly.else30, %polly.stmt.loop_exit.iz
  %polly.indvar_next25 = add nsw i64 %polly.indvar24, 1
  %polly.loop_cond26 = icmp sle i64 %polly.indvar_next25, 5
  br i1 %polly.loop_cond26, label %polly.loop_header21, label %polly.loop_exit23

polly.loop_preheader22:                           ; preds = %polly.stmt.loop_body.k0
  %71 = add i64 %63, %polly.indvar18
  %72 = mul i64 %polly.indvar18, 640
  %scevgep57 = getelementptr i8, i8* %scevgep, i64 %72
  br label %polly.loop_header21

polly.then29:                                     ; preds = %polly.cond27
  br label %polly.stmt.loop_body.k1

polly.stmt.loop_body.k1:                          ; preds = %polly.then29
  %convolution_sum_address.1.s2a.reload = load float, float* %convolution_sum_address.1.s2a
  store float %convolution_sum_address.1.s2a.reload, float* %convolution_sum_address.2.phiops
  br label %polly.loop_preheader32

polly.loop_exit33:                                ; preds = %polly.merge38
  br label %polly.stmt.loop_exit.iz

polly.stmt.loop_exit.iz:                          ; preds = %polly.loop_exit33
  %convolution_sum_address.2.s2a.reload63 = load float, float* %convolution_sum_address.2.s2a
  store float %convolution_sum_address.2.s2a.reload63, float* %convolution_sum_address.1.phiops
  br label %polly.merge28

polly.else30:                                     ; preds = %polly.cond27
  br label %polly.merge28

polly.loop_header31:                              ; preds = %polly.merge38, %polly.loop_preheader32
  %polly.indvar34 = phi i64 [ 0, %polly.loop_preheader32 ], [ %polly.indvar_next35, %polly.merge38 ]
  br label %polly.stmt.loop_header.iz

polly.stmt.loop_header.iz:                        ; preds = %polly.loop_header31
  %convolution_sum_address.2.phiops.reload = load float, float* %convolution_sum_address.2.phiops
  store float %convolution_sum_address.2.phiops.reload, float* %convolution_sum_address.2.s2a
  br label %polly.cond37

polly.cond37:                                     ; preds = %polly.stmt.loop_header.iz
  %73 = icmp eq i64 %polly.indvar34, 0
  br i1 %73, label %polly.then39, label %polly.else40

polly.merge38:                                    ; preds = %polly.else40, %polly.stmt.in-bounds-after
  %polly.indvar_next35 = add nsw i64 %polly.indvar34, 1
  %polly.loop_cond36 = icmp sle i64 %polly.indvar_next35, 1
  br i1 %polly.loop_cond36, label %polly.loop_header31, label %polly.loop_exit33

polly.loop_preheader32:                           ; preds = %polly.stmt.loop_body.k1
  %74 = add i64 %66, %polly.indvar24
  %75 = shl i64 %polly.indvar24, 7
  %scevgep58 = getelementptr i8, i8* %scevgep57, i64 %75
  %scevgep5859 = bitcast i8* %scevgep58 to float*
  br label %polly.loop_header31

polly.then39:                                     ; preds = %polly.cond37
  br label %polly.stmt.loop_body.iz

polly.stmt.loop_body.iz:                          ; preds = %polly.then39
  %p_49 = sdiv i64 %71, 1
  %p_50 = sdiv i64 %74, 1
  store i64 %p_49, i64* %.s2a
  store i64 %p_50, i64* %.s2a51
  br label %polly.cond52

polly.cond52:                                     ; preds = %polly.stmt.loop_body.iz
  %76 = add nsw i64 %polly.indvar, %polly.indvar18
  %77 = icmp sge i64 %76, 30
  %78 = add nsw i64 %polly.indvar, %polly.indvar18
  %79 = icmp sle i64 %78, 1
  %80 = or i1 %77, %79
  %81 = add nsw i64 %polly.indvar6, %polly.indvar24
  %82 = icmp sge i64 %81, 30
  %83 = or i1 %80, %82
  %84 = add nsw i64 %polly.indvar6, %polly.indvar24
  %85 = icmp sle i64 %84, 1
  %86 = or i1 %83, %85
  br i1 %86, label %polly.then54, label %polly.else55

polly.merge53:                                    ; preds = %polly.stmt.in-bounds-true, %polly.stmt.in-bounds-false
  br label %polly.stmt.in-bounds-after

polly.stmt.in-bounds-after:                       ; preds = %polly.merge53
  %convolution_sum_address.3.phiops.reload = load float, float* %convolution_sum_address.3.phiops
  store float %convolution_sum_address.3.phiops.reload, float* %convolution_sum_address.2.phiops
  br label %polly.merge38

polly.else40:                                     ; preds = %polly.cond37
  br label %polly.merge38

polly.then54:                                     ; preds = %polly.cond52
  br label %polly.stmt.in-bounds-false

polly.stmt.in-bounds-false:                       ; preds = %polly.then54
  %convolution_sum_address.2.s2a.reload = load float, float* %convolution_sum_address.2.s2a
  store float %convolution_sum_address.2.s2a.reload, float* %convolution_sum_address.3.phiops
  br label %polly.merge53

polly.else55:                                     ; preds = %polly.cond52
  br label %polly.stmt.in-bounds-true

polly.stmt.in-bounds-true:                        ; preds = %polly.else55
  %.s2a.reload = load i64, i64* %.s2a
  %.s2a51.reload = load i64, i64* %.s2a51
  %convolution_sum_address.2.s2a.reload56 = load float, float* %convolution_sum_address.2.s2a
  %_p_scalar_ = load float, float* %scevgep5859
  %p_ = getelementptr inbounds [1 x [28 x [28 x [1 x float]]]], [1 x [28 x [28 x [1 x float]]]]* @0, i64 0, i64 0, i64 %.s2a.reload, i64 %.s2a51.reload, i64 0
  %_p_scalar_60 = load float, float* %p_
  %p_61 = fmul fast float %_p_scalar_60, %_p_scalar_
  %p_62 = fadd fast float %convolution_sum_address.2.s2a.reload56, %p_61
  store float %p_62, float* %convolution_sum_address.3.phiops
  br label %polly.merge53

polly.loop_header70:                              ; preds = %polly.loop_exit78, %polly.loop_preheader71
  %polly.indvar73 = phi i64 [ 0, %polly.loop_preheader71 ], [ %polly.indvar_next74, %polly.loop_exit78 ]
  br label %polly.loop_preheader77

polly.loop_exit78:                                ; preds = %polly.loop_exit84
  %polly.indvar_next74 = add nsw i64 %polly.indvar73, 1
  %polly.loop_cond75 = icmp sle i64 %polly.indvar_next74, 27
  br i1 %polly.loop_cond75, label %polly.loop_header70, label %polly.loop_exit72

polly.loop_preheader71:                           ; preds = %polly.loop_exit
  br label %polly.loop_header70

polly.loop_header76:                              ; preds = %polly.loop_exit84, %polly.loop_preheader77
  %polly.indvar79 = phi i64 [ 0, %polly.loop_preheader77 ], [ %polly.indvar_next80, %polly.loop_exit84 ]
  br label %polly.loop_preheader83

polly.loop_exit84:                                ; preds = %polly.stmt.loop_body.dim.327
  %polly.indvar_next80 = add nsw i64 %polly.indvar79, 1
  %polly.loop_cond81 = icmp sle i64 %polly.indvar_next80, 27
  br i1 %polly.loop_cond81, label %polly.loop_header76, label %polly.loop_exit78

polly.loop_preheader77:                           ; preds = %polly.loop_header70
  %87 = mul i64 %polly.indvar73, 3584
  %scevgep91 = getelementptr i8, i8* %polly.access.temps.load, i64 %87
  br label %polly.loop_header76

polly.loop_header82:                              ; preds = %polly.stmt.loop_body.dim.327, %polly.loop_preheader83
  %polly.indvar85 = phi i64 [ 0, %polly.loop_preheader83 ], [ %polly.indvar_next86, %polly.stmt.loop_body.dim.327 ]
  br label %polly.stmt.loop_body.dim.327

polly.stmt.loop_body.dim.327:                     ; preds = %polly.loop_header82
  %88 = shl i64 %polly.indvar85, 2
  %scevgep88 = getelementptr i8, i8* %polly.access.params.load, i64 %88
  %scevgep8889 = bitcast i8* %scevgep88 to float*
  %_p_scalar_90 = load float, float* %scevgep8889
  %scevgep93 = getelementptr i8, i8* %scevgep92, i64 %88
  %scevgep9394 = bitcast i8* %scevgep93 to float*
  store float %_p_scalar_90, float* %scevgep9394
  %polly.indvar_next86 = add nsw i64 %polly.indvar85, 1
  %polly.loop_cond87 = icmp sle i64 %polly.indvar_next86, 31
  br i1 %polly.loop_cond87, label %polly.loop_header82, label %polly.loop_exit84

polly.loop_preheader83:                           ; preds = %polly.loop_header76
  %89 = shl i64 %polly.indvar79, 7
  %scevgep92 = getelementptr i8, i8* %scevgep91, i64 %89
  br label %polly.loop_header82

polly.loop_header95:                              ; preds = %polly.loop_exit103, %polly.loop_preheader96
  %polly.indvar98 = phi i64 [ 0, %polly.loop_preheader96 ], [ %polly.indvar_next99, %polly.loop_exit103 ]
  br label %polly.loop_preheader102

polly.loop_exit103:                               ; preds = %polly.loop_exit109
  %polly.indvar_next99 = add nsw i64 %polly.indvar98, 1
  %polly.loop_cond100 = icmp sle i64 %polly.indvar_next99, 27
  br i1 %polly.loop_cond100, label %polly.loop_header95, label %polly.loop_exit97

polly.loop_preheader96:                           ; preds = %polly.loop_exit72
  br label %polly.loop_header95

polly.loop_header101:                             ; preds = %polly.loop_exit109, %polly.loop_preheader102
  %polly.indvar104 = phi i64 [ 0, %polly.loop_preheader102 ], [ %polly.indvar_next105, %polly.loop_exit109 ]
  br label %polly.loop_preheader108

polly.loop_exit109:                               ; preds = %polly.stmt.loop_body.dim.351
  %polly.indvar_next105 = add nsw i64 %polly.indvar104, 1
  %polly.loop_cond106 = icmp sle i64 %polly.indvar_next105, 27
  br i1 %polly.loop_cond106, label %polly.loop_header101, label %polly.loop_exit103

polly.loop_preheader102:                          ; preds = %polly.loop_header95
  %90 = mul i64 %polly.indvar98, 3584
  %scevgep113 = getelementptr i8, i8* %4, i64 %90
  %scevgep118 = getelementptr i8, i8* %polly.access.temps.load, i64 %90
  %scevgep124 = getelementptr i8, i8* %retval, i64 %90
  br label %polly.loop_header101

polly.loop_header107:                             ; preds = %polly.stmt.loop_body.dim.351, %polly.loop_preheader108
  %polly.indvar110 = phi i64 [ 0, %polly.loop_preheader108 ], [ %polly.indvar_next111, %polly.stmt.loop_body.dim.351 ]
  br label %polly.stmt.loop_body.dim.351

polly.stmt.loop_body.dim.351:                     ; preds = %polly.loop_header107
  %91 = shl i64 %polly.indvar110, 2
  %scevgep115 = getelementptr i8, i8* %scevgep114, i64 %91
  %scevgep115116 = bitcast i8* %scevgep115 to float*
  %_p_scalar_117 = load float, float* %scevgep115116
  %scevgep120 = getelementptr i8, i8* %scevgep119, i64 %91
  %scevgep120121 = bitcast i8* %scevgep120 to float*
  %_p_scalar_122 = load float, float* %scevgep120121
  %p_123 = fadd fast float %_p_scalar_117, %_p_scalar_122
  %92 = shl i64 %polly.indvar110, 2
  %scevgep126 = getelementptr i8, i8* %scevgep125, i64 %92
  %scevgep126127 = bitcast i8* %scevgep126 to float*
  store float %p_123, float* %scevgep126127
  %polly.indvar_next111 = add nsw i64 %polly.indvar110, 1
  %polly.loop_cond112 = icmp sle i64 %polly.indvar_next111, 31
  br i1 %polly.loop_cond112, label %polly.loop_header107, label %polly.loop_exit109

polly.loop_preheader108:                          ; preds = %polly.loop_header101
  %93 = shl i64 %polly.indvar104, 7
  %scevgep114 = getelementptr i8, i8* %scevgep113, i64 %93
  %scevgep119 = getelementptr i8, i8* %scevgep118, i64 %93
  %scevgep125 = getelementptr i8, i8* %scevgep124, i64 %93
  br label %polly.loop_header107
}

attributes #0 = { "polly-optimized" }

!0 = !{!1, !1, i64 0}
!1 = !{!"pointer-to element_type: F32 dimensions: 5 dimensions: 5 dimensions: 1 dimensions: 32 layout { minor_to_major: 3 minor_to_major: 2 minor_to_major: 1 minor_to_major: 0 }", !2}
!2 = !{!"XLA TBAA"}
!3 = !{i64 3200}
!4 = !{i64 16}
!5 = !{!6, !6, i64 0}
!6 = !{!"pointer-to element_type: F32 dimensions: 1 dimensions: 28 dimensions: 28 dimensions: 32 layout { minor_to_major: 3 minor_to_major: 2 minor_to_major: 1 minor_to_major: 0 }", !2}
!7 = !{}
!8 = !{i64 100352}
!9 = !{!10, !10, i64 0}
!10 = !{!"element_type: F32 dimensions: 1 dimensions: 28 dimensions: 28 dimensions: 32 layout { minor_to_major: 3 minor_to_major: 2 minor_to_major: 1 minor_to_major: 0 }", !2}
!11 = !{!12}
!12 = !{!"buffer: 0", !13}
!13 = distinct !{!13}
!14 = !{!15}
!15 = !{!"buffer: 1", !13}
!16 = !{!17, !17, i64 0}
!17 = !{!"pointer-to element_type: F32 dimensions: 32 layout { minor_to_major: 0 }", !2}
!18 = !{i64 128}
!19 = !{!20, !20, i64 0}
!20 = !{!"element_type: F32 dimensions: 5 dimensions: 5 dimensions: 1 dimensions: 32 layout { minor_to_major: 3 minor_to_major: 2 minor_to_major: 1 minor_to_major: 0 }", !2}
!21 = !{!22, !22, i64 0}
!22 = !{!"element_type: F32 dimensions: 1 dimensions: 28 dimensions: 28 dimensions: 1 layout { minor_to_major: 3 minor_to_major: 2 minor_to_major: 1 minor_to_major: 0 }", !2}
!23 = !{!24, !24, i64 0}
!24 = !{!"element_type: F32 dimensions: 32 layout { minor_to_major: 0 }", !2}