Example input and output from Julia using LLVM 3.6.1 and JULIA_LLVM_ARGS=-debug-only=loop-vectorize

input.jl

function mysum!{T}(F, A::AbstractArray{T})
    for j = 1:size(A,2)
        @simd for i = 1:size(A,1)
            @inbounds F[i,j] += A[i,j]
        end
    end
    return F
end

output.txt

LV: Checking a loop in "julia_mysum!_20989" from /localdisk/adrobiso/julia-trunk/b.cpp:75:1
LV: Loop hints: force=? width=0 unroll=0
LV: Found a loop: L14
LV: Found an induction variable.
LV: A loop annotated parallel, ignore memory dependency checks.
LV: We can vectorize this loop!
LV: Found trip count: 0
LV: The Widest type: 64 bits.
LV: The Widest register is: 256 bits.
LV: Found an estimated cost of 0 for VF 1 For instruction:   %"##i#1682.0" = phi i64 [ 0, %if12 ], [ %56, %L14 ]
LV: Found an estimated cost of 1 for VF 1 For instruction:   %37 = add i64 %"##i#1682.0", %35, !dbg !35
LV: Found an estimated cost of 0 for VF 1 For instruction:   %38 = getelementptr float* %11, i64 %37, !dbg !35
LV: Found an estimated cost of 1 for VF 1 For instruction:   %39 = load float* %38, align 4, !dbg !35, !tbaa !44, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 1 for VF 1 For instruction:   %40 = load %jl_value_t** %17, align 8, !dbg !35, !tbaa !47, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 1 for VF 1 For instruction:   %41 = load i64* %19, align 8, !dbg !35, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 1 for VF 1 For instruction:   %42 = load i64* %21, align 8, !dbg !35, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 0 for VF 1 For instruction:   %43 = getelementptr inbounds %jl_value_t* %40, i64 3, i32 0, !dbg !35
LV: Found an estimated cost of 0 for VF 1 For instruction:   %44 = bitcast %jl_value_t** %43 to i64*, !dbg !35
LV: Found an estimated cost of 1 for VF 1 For instruction:   %45 = load i64* %44, align 8, !dbg !35, !tbaa !31, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 1 for VF 1 For instruction:   %46 = add i64 %36, %42, !dbg !35
LV: Found an estimated cost of 1 for VF 1 For instruction:   %47 = mul i64 %46, %45, !dbg !35
LV: Found an estimated cost of 1 for VF 1 For instruction:   %48 = add nsw i64 %"##i#1682.0", -1, !dbg !35
LV: Found an estimated cost of 1 for VF 1 For instruction:   %49 = add i64 %48, %41, !dbg !35
LV: Found an estimated cost of 1 for VF 1 For instruction:   %50 = add i64 %49, %47, !dbg !35
LV: Found an estimated cost of 0 for VF 1 For instruction:   %51 = bitcast %jl_value_t* %40 to float**, !dbg !35
LV: Found an estimated cost of 1 for VF 1 For instruction:   %52 = load float** %51, align 8, !dbg !35, !tbaa !27, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 0 for VF 1 For instruction:   %53 = getelementptr float* %52, i64 %50, !dbg !35
LV: Found an estimated cost of 1 for VF 1 For instruction:   %54 = load float* %53, align 4, !dbg !35, !tbaa !44, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 2 for VF 1 For instruction:   %55 = fadd float %39, %54, !dbg !35
LV: Found an estimated cost of 1 for VF 1 For instruction:   store float %55, float* %38, align 4, !dbg !35, !tbaa !44, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 1 for VF 1 For instruction:   %56 = add nuw nsw i64 %"##i#1682.0", 1, !dbg !48, !simd_loop !4
LV: Found an estimated cost of 1 for VF 1 For instruction:   %exitcond = icmp eq i64 %56, %32, !dbg !49
LV: Found an estimated cost of 0 for VF 1 For instruction:   br i1 %exitcond, label %L21.loopexit, label %L14, !dbg !49, !llvm.loop !46
LV: Scalar loop costs: 18.
LV: Found an estimated cost of 0 for VF 2 For instruction:   %"##i#1682.0" = phi i64 [ 0, %if12 ], [ %56, %L14 ]
LV: Found an estimated cost of 1 for VF 2 For instruction:   %37 = add i64 %"##i#1682.0", %35, !dbg !35
LV: Found an estimated cost of 0 for VF 2 For instruction:   %38 = getelementptr float* %11, i64 %37, !dbg !35
LV: Found an estimated cost of 1 for VF 2 For instruction:   %39 = load float* %38, align 4, !dbg !35, !tbaa !44, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 26 for VF 2 For instruction:   %40 = load %jl_value_t** %17, align 8, !dbg !35, !tbaa !47, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 26 for VF 2 For instruction:   %41 = load i64* %19, align 8, !dbg !35, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 26 for VF 2 For instruction:   %42 = load i64* %21, align 8, !dbg !35, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 0 for VF 2 For instruction:   %43 = getelementptr inbounds %jl_value_t* %40, i64 3, i32 0, !dbg !35
LV: Found an estimated cost of 0 for VF 2 For instruction:   %44 = bitcast %jl_value_t** %43 to i64*, !dbg !35
LV: Found an estimated cost of 26 for VF 2 For instruction:   %45 = load i64* %44, align 8, !dbg !35, !tbaa !31, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 1 for VF 2 For instruction:   %46 = add i64 %36, %42, !dbg !35
LV: Found an estimated cost of 9 for VF 2 For instruction:   %47 = mul i64 %46, %45, !dbg !35
LV: Found an estimated cost of 1 for VF 2 For instruction:   %48 = add nsw i64 %"##i#1682.0", -1, !dbg !35
LV: Found an estimated cost of 1 for VF 2 For instruction:   %49 = add i64 %48, %41, !dbg !35
LV: Found an estimated cost of 1 for VF 2 For instruction:   %50 = add i64 %49, %47, !dbg !35
LV: Found an estimated cost of 0 for VF 2 For instruction:   %51 = bitcast %jl_value_t* %40 to float**, !dbg !35
LV: Found an estimated cost of 26 for VF 2 For instruction:   %52 = load float** %51, align 8, !dbg !35, !tbaa !27, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 0 for VF 2 For instruction:   %53 = getelementptr float* %52, i64 %50, !dbg !35
LV: Found an estimated cost of 25 for VF 2 For instruction:   %54 = load float* %53, align 4, !dbg !35, !tbaa !44, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 2 for VF 2 For instruction:   %55 = fadd float %39, %54, !dbg !35
LV: Found an estimated cost of 1 for VF 2 For instruction:   store float %55, float* %38, align 4, !dbg !35, !tbaa !44, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 1 for VF 2 For instruction:   %56 = add nuw nsw i64 %"##i#1682.0", 1, !dbg !48, !simd_loop !4
LV: Found an estimated cost of 1 for VF 2 For instruction:   %exitcond = icmp eq i64 %56, %32, !dbg !49
LV: Found an estimated cost of 0 for VF 2 For instruction:   br i1 %exitcond, label %L21.loopexit, label %L14, !dbg !49, !llvm.loop !46
LV: Vector loop of width 2 costs: 87.
LV: Found an estimated cost of 0 for VF 4 For instruction:   %"##i#1682.0" = phi i64 [ 0, %if12 ], [ %56, %L14 ]
LV: Found an estimated cost of 1 for VF 4 For instruction:   %37 = add i64 %"##i#1682.0", %35, !dbg !35
LV: Found an estimated cost of 0 for VF 4 For instruction:   %38 = getelementptr float* %11, i64 %37, !dbg !35
LV: Found an estimated cost of 1 for VF 4 For instruction:   %39 = load float* %38, align 4, !dbg !35, !tbaa !44, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 52 for VF 4 For instruction:   %40 = load %jl_value_t** %17, align 8, !dbg !35, !tbaa !47, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 52 for VF 4 For instruction:   %41 = load i64* %19, align 8, !dbg !35, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 52 for VF 4 For instruction:   %42 = load i64* %21, align 8, !dbg !35, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 0 for VF 4 For instruction:   %43 = getelementptr inbounds %jl_value_t* %40, i64 3, i32 0, !dbg !35
LV: Found an estimated cost of 0 for VF 4 For instruction:   %44 = bitcast %jl_value_t** %43 to i64*, !dbg !35
LV: Found an estimated cost of 52 for VF 4 For instruction:   %45 = load i64* %44, align 8, !dbg !35, !tbaa !31, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 1 for VF 4 For instruction:   %46 = add i64 %36, %42, !dbg !35
LV: Found an estimated cost of 9 for VF 4 For instruction:   %47 = mul i64 %46, %45, !dbg !35
LV: Found an estimated cost of 1 for VF 4 For instruction:   %48 = add nsw i64 %"##i#1682.0", -1, !dbg !35
LV: Found an estimated cost of 1 for VF 4 For instruction:   %49 = add i64 %48, %41, !dbg !35
LV: Found an estimated cost of 1 for VF 4 For instruction:   %50 = add i64 %49, %47, !dbg !35
LV: Found an estimated cost of 0 for VF 4 For instruction:   %51 = bitcast %jl_value_t* %40 to float**, !dbg !35
LV: Found an estimated cost of 52 for VF 4 For instruction:   %52 = load float** %51, align 8, !dbg !35, !tbaa !27, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 0 for VF 4 For instruction:   %53 = getelementptr float* %52, i64 %50, !dbg !35
LV: Found an estimated cost of 51 for VF 4 For instruction:   %54 = load float* %53, align 4, !dbg !35, !tbaa !44, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 2 for VF 4 For instruction:   %55 = fadd float %39, %54, !dbg !35
LV: Found an estimated cost of 1 for VF 4 For instruction:   store float %55, float* %38, align 4, !dbg !35, !tbaa !44, !llvm.mem.parallel_loop_access !45
LV: Found an estimated cost of 1 for VF 4 For instruction:   %56 = add nuw nsw i64 %"##i#1682.0", 1, !dbg !48, !simd_loop !4
LV: Found an estimated cost of 1 for VF 4 For instruction:   %exitcond = icmp eq i64 %56, %32, !dbg !49
LV: Found an estimated cost of 0 for VF 4 For instruction:   br i1 %exitcond, label %L21.loopexit, label %L14, !dbg !49, !llvm.loop !46
LV: Vector loop of width 4 costs: 82.
LV: Selecting VF: 1.
LV: The target has 16 registers
LV(REG): Calculating max register usage:
LV(REG): At #0 Interval # 0
LV(REG): At #1 Interval # 1
LV(REG): At #2 Interval # 2
LV(REG): At #3 Interval # 2
LV(REG): At #4 Interval # 3
LV(REG): At #5 Interval # 4
LV(REG): At #6 Interval # 5
LV(REG): At #7 Interval # 6
LV(REG): At #8 Interval # 7
LV(REG): At #9 Interval # 7
LV(REG): At #10 Interval # 7
LV(REG): At #11 Interval # 7
LV(REG): At #12 Interval # 6
LV(REG): At #13 Interval # 7
LV(REG): At #14 Interval # 6
LV(REG): At #15 Interval # 5
LV(REG): At #16 Interval # 5
LV(REG): At #17 Interval # 5
LV(REG): At #18 Interval # 4
LV(REG): At #19 Interval # 4
LV(REG): At #21 Interval # 3
LV(REG): At #23 Interval # 4
LV(REG): Found max usage: 7
LV(REG): Found invariant usage: 7
LV(REG): LoopSize: 25
LV: Loop cost is 18
LV: Unrolling to reduce branch cost.
LV: Found a vectorizable loop (1) in /localdisk/adrobiso/julia-trunk/b.cpp:75:1
LV: Unroll Factor is 1
LV: Vectorization is possible but not beneficial

; Function Attrs: sspreq
define %jl_value_t* @"julia_mysum!_20989"(%jl_value_t*, %jl_value_t**, i32) #-1 {
top:
  %3 = load %jl_value_t** %1, align 8
  %4 = getelementptr %jl_value_t** %1, i64 1
  %5 = load %jl_value_t** %4, align 8
  %6 = getelementptr inbounds %jl_value_t* %5, i64 6
  %7 = bitcast %jl_value_t* %6 to i64*
  %8 = load i64* %7, align 8
  %9 = icmp sgt i64 %8, 0
  br i1 %9, label %L3.preheader, label %L23

L3.preheader:                                     ; preds = %top
  %10 = bitcast %jl_value_t* %3 to float**
  %11 = load float** %10, align 8
  %12 = getelementptr inbounds %jl_value_t* %3, i64 3, i32 0
  %13 = bitcast %jl_value_t** %12 to i64*
  %14 = load i64* %13, align 8
  %15 = getelementptr %jl_value_t* %5, i64 5
  %16 = bitcast %jl_value_t* %15 to i64*
  %17 = getelementptr inbounds %jl_value_t* %5, i64 0, i32 0
  %18 = getelementptr %jl_value_t* %5, i64 1
  %19 = bitcast %jl_value_t* %18 to i64*
  %20 = getelementptr inbounds %jl_value_t* %5, i64 3
  %21 = bitcast %jl_value_t* %20 to i64*
  br label %L3

L3:                                               ; preds = %L21, %L3.preheader
  %"#s1.0" = phi i64 [ %57, %L21 ], [ 1, %L3.preheader ]
  %22 = load i64* %16, align 8
  %23 = icmp sgt i64 %22, 0
  %24 = select i1 %23, i64 %22, i64 0
  %25 = call { i64, i1 } @llvm.ssub.with.overflow.i64(i64 %24, i64 1)
  %26 = extractvalue { i64, i1 } %25, 1
  br i1 %26, label %fail.split, label %L3.L3.split_crit_edge

L3.L3.split_crit_edge:                            ; preds = %L3
  %27 = extractvalue { i64, i1 } %25, 0
  %28 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %27, i64 1)
  %29 = extractvalue { i64, i1 } %28, 1
  br i1 %29, label %fail10, label %pass11

fail.split:                                       ; preds = %L3
  %30 = load %jl_value_t** @jl_overflow_exception, align 8
  call void @jl_throw_with_superfluous_argument(%jl_value_t* %30, i32 67)
  unreachable

fail10:                                           ; preds = %L3.L3.split_crit_edge
  %31 = load %jl_value_t** @jl_overflow_exception, align 8
  call void @jl_throw_with_superfluous_argument(%jl_value_t* %31, i32 67)
  unreachable

pass11:                                           ; preds = %L3.L3.split_crit_edge
  %32 = extractvalue { i64, i1 } %28, 0
  %33 = icmp slt i64 %32, 1
  br i1 %33, label %L21, label %if12

if12:                                             ; preds = %pass11
  %34 = add i64 %"#s1.0", -1
  %35 = mul i64 %34, %14
  %36 = add i64 %"#s1.0", -2
  br label %L14

L14:                                              ; preds = %L14, %if12
  %"##i#1682.0" = phi i64 [ 0, %if12 ], [ %56, %L14 ]
  %37 = add i64 %"##i#1682.0", %35
  %38 = getelementptr float* %11, i64 %37
  %39 = load float* %38, align 4
  %40 = load %jl_value_t** %17, align 8
  %41 = load i64* %19, align 8
  %42 = load i64* %21, align 8
  %43 = getelementptr inbounds %jl_value_t* %40, i64 3, i32 0
  %44 = bitcast %jl_value_t** %43 to i64*
  %45 = load i64* %44, align 8
  %46 = add i64 %36, %42
  %47 = mul i64 %46, %45
  %48 = add nsw i64 %"##i#1682.0", -1
  %49 = add i64 %48, %41
  %50 = add i64 %49, %47
  %51 = bitcast %jl_value_t* %40 to float**
  %52 = load float** %51, align 8
  %53 = getelementptr float* %52, i64 %50
  %54 = load float* %53, align 4
  %55 = fadd float %39, %54
  store float %55, float* %38, align 4
  %56 = add nuw nsw i64 %"##i#1682.0", 1
  %exitcond = icmp eq i64 %56, %32
  br i1 %exitcond, label %L21.loopexit, label %L14

L21.loopexit:                                     ; preds = %L14
  br label %L21

L21:                                              ; preds = %L21.loopexit, %pass11
  %57 = add i64 %"#s1.0", 1
  %58 = icmp eq i64 %"#s1.0", %8
  br i1 %58, label %L23.loopexit, label %L3

L23.loopexit:                                     ; preds = %L21
  br label %L23

L23:                                              ; preds = %L23.loopexit, %top
  ret %jl_value_t* %3
}