Use multi-arg pick_vecotor_width val function with _avx_!.

Author: chriselrod

src/lowering.jl

@@ -344,8 +344,10 @@ function setup_preamble!(ls::LoopSet, us::UnrollSpecification)
     vectorized = order[vectorizedloopnum]
     # println("Setup preamble")
     W = ls.W; typeT = ls.T
-    length(ls.includedarrays) == 0 || push!(ls.preamble.args, Expr(:(=), typeT, determine_eltype(ls)))
-    push!(ls.preamble.args, Expr(:(=), W, determine_width(ls, vectorized)))
+    if length(ls.includedarrays) >= 0
+        push!(ls.preamble.args, Expr(:(=), typeT, determine_eltype(ls)))
+        push!(ls.preamble.args, Expr(:(=), W, determine_width(ls, vectorized)))
+    end
     lower_licm_constants!(ls)
     pushpreamble!(ls, definemask(getloop(ls, vectorized), W))#, U > 1 && unrolledloopnum == vectorizedloopnum))
     for op ∈ operations(ls)

src/reconstruct_loopset.jl

@@ -383,6 +383,7 @@ function avx_loopset(instr, ops, arf, AM, LPSYM, LB, vargs)
     expandedv = [isexpanded(ls, ops, nopsv, i) for i ∈ eachindex(ops)]
     mrefs = create_mrefs!(ls, arf, arraysymbolinds, opsymbols, nopsv, expandedv, vargs)
     pushpreamble!(ls, Expr(:(=), ls.T, Expr(:call, :promote_type, [Expr(:call, :eltype, vptr(mref)) for mref ∈ mrefs]...)))
+    pushpreamble!(ls, Expr(:(=), ls.W, Expr(:call, lv(:pick_vector_width_val), [Expr(:call, :eltype, vptr(mref)) for mref ∈ mrefs]...)))
     num_params = num_arrays + num_parameters(AM)
     num_params = add_ops!(ls, instr, ops, mrefs, opsymbols, num_params, nopsv, expandedv, elementbytes)
     process_metadata!(ls, AM, length(arf))

test/ifelsemasks.jl

@@ -68,7 +68,6 @@ T = Float32
         t
     end
 
-
     function addormul!(c, a, b)
         for i ∈ eachindex(c,a,b)
             c[i] = a[i] > b[i] ? a[i] + b[i] : a[i] * b[i]