Remove println statements accompanying errors

[mcabbott]

This replaces println(ex); throw(str) with an error type, so that if macro expansion fails within a try-catch block, nothing is printed.

Closes mcabbott/Tullio.jl#26

Example of new behaviour:

julia> try                                                                   
       macroexpand(Main, quote                                               
       @avx for i = 𝒶𝓍i                                                      
           if i ∈ (axes)(var"𝛥≪rand(10)≫", 1)                                
               var"𝛥≪rand(10)≫"[i] = var"𝛥≪rand(10)≫"[i] + 𝛥ℛ[i]             
           else                                                              
               (zero)((eltype)(var"𝛥≪rand(10)≫"))                            
           end                                                               
       end                                                                   
       end)                                                                  
       catch err                                                             
       end  # silence!

julia> @macroexpand @avx for i = 𝒶𝓍i                                                      
           if i ∈ (axes)(var"𝛥≪rand(10)≫", 1)                                
               var"𝛥≪rand(10)≫"[i] = var"𝛥≪rand(10)≫"[i] + 𝛥ℛ[i]             
           else                                                              
               (zero)((eltype)(var"𝛥≪rand(10)≫"))                            
           end                                                               
       end  # similar output to before:
ERROR: LoadError: Don't know how to handle expression.
if i ∈ axes(var"𝛥≪rand(10)≫", 1)
    #= REPL[10]:3 =#
    var"𝛥≪rand(10)≫"[i] = var"𝛥≪rand(10)≫"[i] + 𝛥ℛ[i]
else
    #= REPL[10]:5 =#
    zero(eltype(var"𝛥≪rand(10)≫"))
end
Stacktrace:
 [1] push!(ls::LoopVectorization.LoopSet, ex::Expr, elementbytes::Int64, position::Int64)
   @ LoopVectorization ~/.julia/dev/LoopVectorization/src/graphs.jl:754
 [2] add_block!
...

[chriselrod]

Looks good to me.
Feel free to bump the patch version.

[mcabbott]

Great, done!

[codecov-io]

Codecov Report

Merging #172 (0ef644f) into master (b156769) will increase coverage by 1.92%.
The diff coverage is 37.50%.

@@            Coverage Diff             @@
##           master     #172      +/-   ##
==========================================
+ Coverage   92.75%   94.67%   +1.92%     
==========================================
  Files          32       32              
  Lines        4516     4511       -5     
==========================================
+ Hits         4189     4271      +82     
+ Misses        327      240      -87     

Impacted Files	Coverage Δ
src/graphs.jl	88.49% <37.50%> (+1.49%)	⬆️
src/lower_memory_common.jl	97.50% <0.00%> (+1.66%)	⬆️
src/lowering.jl	97.63% <0.00%> (+2.36%)	⬆️
src/determinestrategy.jl	99.84% <0.00%> (+2.39%)	⬆️
src/reconstruct_loopset.jl	96.79% <0.00%> (+6.04%)	⬆️
src/lower_load.jl	98.36% <0.00%> (+19.67%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update b156769...0ef644f. Read the comment docs.

[chriselrod]

I'll merge and issue another release once tests pass.

[chriselrod]

Perhaps add a

@test_throws LoadError @macroexpand @avx for i = 𝒶𝓍i                                                      
    if i ∈ (axes)(var"𝛥≪rand(10)≫", 1)                                
        var"𝛥≪rand(10)≫"[i] = var"𝛥≪rand(10)≫"[i] + 𝛥ℛ[i]             
    else                                                              
        (zero)((eltype)(var"𝛥≪rand(10)≫"))                            
    end                                                               
end

?

[mcabbott]

Done. Assuming that this kind of if-else is something unlikely to ever work, I guess. It was just the first one I happened to pick up for quick tests locally.

By the way, there isn't an AVX-friendly way of doing that, is there? I mean of testing some index & returning zero (or some other padding) out of bounds? Something like ifelse(i in axes(ys,1), ys[i], 0.0) will actually read out of bounds

[chriselrod]

Yeah, maybe it's better to have a wonkier example now, but I figured it wouldn't be that bad to change it later.

By the way, there isn't an AVX-friendly way of doing that, is there? I mean of testing some index & returning zero (or some other padding) out of bounds? Something like ifelse(i in axes(ys,1), ys[i], 0.0) will actually read out of bounds

There is, see here for an example.
If you use ? instead of ifelse, LoopVectorization will convert that into a masked load.
So you'd want to do something similar to the linked line.
Instead of

xim = i > ifirst ? x[Ipre, i-1, Ipost] : xi

you'd want

((i >= first(axes(ys,1))) & (i <= last(axes(ys,1)))) ? ys[i] : 0.0

and that should work already.

AVX has a masked move instruction:
https://www.felixcloutier.com/x86/vmaskmov
While with AVX512, it of course uses a regular move but applies a mask.
These don't read out of bounds and won't trigger segfaults.

[mcabbott]

That looks great, thanks! Will try this out.