track groupedstridedpointer returning tuple to GC.@preserve, and use comparison operators for AbstractStridedPointers now that they're available instead of pointerforcomparison.

Author: chriselrod

src/LoopVectorization.jl

@@ -6,7 +6,7 @@ using VectorizationBase: register_size, register_count, cache_linesize, cache_si
     mask, pick_vector_width, MM, AbstractMask, data, grouped_strided_pointer,
     maybestaticlength, maybestaticsize, staticm1, staticp1, staticmul, vzero,
     maybestaticrange, offsetprecalc, lazymul,
-    maybestaticfirst, maybestaticlast, scalar_less, scalar_greaterequal, gep, gesp, llvmptr, NativeTypes,
+    maybestaticfirst, maybestaticlast, scalar_less, scalar_greaterequal, gep, gesp, NativeTypes, #llvmptr,
     vfmadd, vfmsub, vfnmadd, vfnmsub, vfmadd_fast, vfmsub_fast, vfnmadd_fast, vfnmsub_fast, vfmadd231, vfmsub231, vfnmadd231, vfnmsub231,
     vfma_fast, vmuladd_fast, vdiv_fast, vadd_fast, vsub_fast, vmul_fast,
     relu, stridedpointer, StridedPointer, StridedBitPointer, AbstractStridedPointer, _vload, _vstore!,

src/broadcast.jl

@@ -207,7 +207,8 @@ end
         Nnew += 1
     end
     typ = Expr(:curly, :StridedPointer, T, Nnew, Cnew, Bnew, Rtup)
-    ptr = Expr(:call, typ, Expr(:call, lv(:llvmptr), :p), strd, offsets)
+    # ptr = Expr(:call, typ, Expr(:call, lv(:llvmptr), :p), strd, offsets)
+    ptr = Expr(:call, typ, Expr(:call, lv(:pointer), :p), strd, offsets)
     Expr(:block, Expr(:meta,:inline), :(strd = p.strd), :(offs = p.offsets), ptr)
 end
 # @generated function VectorizationBase.stridedpointer(A::LowDimArray{D,T,N}) where {D,T,N}

src/codegen/loopstartstopmanager.jl

@@ -65,7 +65,8 @@ end
     newB = C > 0 ? (C == minrank ? B : 0) : B #TODO: confirm correctness
     quote
         $(Expr(:meta,:inline))
-        VectorizationBase.StridedPointer{$T,1,$newC,$newB,$(R[minrank],)}($(lv(llvmptr))(sptr), (sptr.strd[$minrank],), (Zero(),))
+        # VectorizationBase.StridedPointer{$T,1,$newC,$newB,$(R[minrank],)}($(lv(llvmptr))(sptr), (sptr.strd[$minrank],), (Zero(),))
+        VectorizationBase.StridedPointer{$T,1,$newC,$newB,$(R[minrank],)}(pointer(sptr), (sptr.strd[$minrank],), (Zero(),))
     end
 end
 set_first_stride(x) = x # cross fingers that this works

src/codegen/lowering.jl

@@ -655,7 +655,7 @@ function init_remblock(unrolledloop::Loop, lssm::LoopStartStopManager, n::Int)#u
     else
         termar = lssm.incrementedptrs[n][termind]
         ptr = vptr(termar)
-        condition = Expr(:call, :<, vptr(ptr), maxsym(ptr, 0))
+        condition = Expr(:call, :<, ptr, maxsym(ptr, 0))
     end
     Expr(:if, condition)
 end

src/condense_loopset.jl

@@ -230,6 +230,7 @@ function add_grouped_strided_pointer!(extra_args::Expr, ls::LoopSet)
     gsp = Expr(:call, lv(:grouped_strided_pointer))
     tgarrays = Expr(:tuple)
     i = 0
+    preserve_assignment = Expr(:tuple); preserve = Symbol[];
     for ref ∈ ls.refs_aliasing_syms
         i += 1
         found = false
@@ -242,6 +243,9 @@ function add_grouped_strided_pointer!(extra_args::Expr, ls::LoopSet)
             break
         end
         found || push!(tgarrays.args, vptr(ref))
+        pres = gensym!(ls, "#preserve#")
+        push!(preserve_assignment.args, pres)
+        push!(preserve, pres)
     end
     push!(gsp.args, tgarrays)
     matcheddims = Expr(:tuple)
@@ -255,8 +259,10 @@ function add_grouped_strided_pointer!(extra_args::Expr, ls::LoopSet)
         length(t.args) > 1 && push!(matcheddims.args, t)
     end
     push!(gsp.args, val(matcheddims))
-    push!(extra_args.args, gsp)
-    nothing
+    gsps = gensym!(ls, "#grouped#strided#pointer#")
+    push!(extra_args.args, gsps)
+    pushpreamble!(ls, Expr(:(=), Expr(:tuple, gsps, preserve_assignment), gsp))
+    preserve
 end
 
 # first_cache() = ifelse(gt(num_cache_levels(), StaticInt{2}()), StaticInt{2}(), StaticInt{1}())
@@ -309,7 +315,7 @@ function generate_call(ls::LoopSet, (inline,u₁,u₂)::Tuple{Bool,Int8,Int8}, t
     vargs_as_tuple = true#!debug
     vargs_as_tuple || push!(q.args, lbarg)
     extra_args = vargs_as_tuple ? Expr(:tuple) : q
-    add_grouped_strided_pointer!(extra_args, ls)
+    preserve = add_grouped_strided_pointer!(extra_args, ls)
     for is ∈ ls.preamble_symsym
         push!(extra_args.args, last(is))
     end
@@ -320,7 +326,7 @@ function generate_call(ls::LoopSet, (inline,u₁,u₂)::Tuple{Bool,Int8,Int8}, t
     vargs_as_tuple && push!(q.args, Expr(:tuple, lbarg, extra_args))
     vecwidthdefq = Expr(:block)
     define_eltype_vec_width!(vecwidthdefq, ls, nothing)
-    Expr(:block, vecwidthdefq, q)
+    Expr(:block, vecwidthdefq, q), preserve
 end
 
 
@@ -383,16 +389,22 @@ make_crashy(q) = Expr(:macrocall, Symbol("@inbounds"), LineNumberNode(@__LINE__,
 @inline vecmemaybe(x::VectorizationBase._Vec) = Vec(x)
 @inline vecmemaybe(x::Tuple) = VectorizationBase.VecUnroll(x)
 
+function gc_preserve(call::Expr, preserve::Vector{Symbol})
+    q = Expr(:gc_preserve, call)
+    append!(q.args, preserve)
+    q
+end
+
 function setup_call_inline(ls::LoopSet, inline::Bool, u₁::Int8, u₂::Int8, thread::Int)
-    call = generate_call(ls, (inline,u₁,u₂), thread % UInt, false)
+    call, preserve = generate_call(ls, (inline,u₁,u₂), thread % UInt, false)
     if iszero(length(ls.outer_reductions))
-        q = Expr(:block,gc_preserve(ls, call))
-        append!(ls.preamble.args, q.args)
+        pushpreamble!(ls, gc_preserve(call, preserve))
+        push!(ls.preamble.args, nothing)
         return ls.preamble
     end
     retv = loopset_return_value(ls, Val(false))
     outer_reducts = Expr(:local)
-    q = Expr(:block,gc_preserve(ls, Expr(:(=), retv, call)))
+    q = Expr(:block,gc_preserve(Expr(:(=), retv, call), preserve))
     for or ∈ ls.outer_reductions
         op = ls.operations[or]
         var = name(op)
@@ -409,7 +421,7 @@ function setup_call_inline(ls::LoopSet, inline::Bool, u₁::Int8, u₂::Int8, th
 end
 function setup_call_debug(ls::LoopSet)
     # avx_loopset(instr, ops, arf, AM, LB, vargs)
-    pushpreamble!(ls, generate_call(ls, (false,zero(Int8),zero(Int8)), zero(UInt), true))
+    pushpreamble!(ls, first(generate_call(ls, (false,zero(Int8),zero(Int8)), zero(UInt), true)))
     Expr(:block, ls.prepreamble, ls.preamble)
 end
 function setup_call(

src/reconstruct_loopset.jl

@@ -171,6 +171,7 @@ function rank_to_sortperm(R::NTuple{N,Int}) where {N}
     end
     sp
 end
+# sptrs::Expr, ls::LoopSet, ar::ArrayReferenceMeta, @nospecialize(_::Type{Core.LLVMPtr{T,0}}),
 function add_mref!(
     sptrs::Expr, ls::LoopSet, ar::ArrayReferenceMeta, @nospecialize(_::Type{Ptr{T}}),
     C::Int, B::Int, R::NTuple{N,Int}, name::Symbol
@@ -425,6 +426,7 @@ end
 # elbytes(::VectorizationBase.AbstractPointer{T}) where {T} = sizeof(T)::Int
 # typeeltype(::Type{P}) where {T,P<:VectorizationBase.AbstractStridedPointer{T}} = T
 typeeltype(::Type{Ptr{T}}) where {T} = T
+# typeeltype(::Type{Core.LLVMPtr{T,0}}) where {T} = T
 typeeltype(::Type{VectorizationBase.FastRange{T,F,S}}) where {T,F,S} = T
 typeeltype(::Type{T}) where {T<:Real} = T
 # typeeltype(::Any) = Int8
@@ -568,7 +570,7 @@ Execute an `@avx` block. The block's code is represented via the arguments:
 @generated function _avx_!(
     ::Val{UNROLL}, ::Val{OPS}, ::Val{ARF}, ::Val{AM}, ::Val{LPSYM}, var"#lv#tuple#args#"::Tuple{LB,V}
 ) where {UNROLL, OPS, ARF, AM, LPSYM, LB, V}
-    # 1 + 1 # Irrelevant line you can comment out/in to force recompilation...
+    1 + 1 # Irrelevant line you can comment out/in to force recompilation...
     ls = _avx_loopset(OPS, ARF, AM, LPSYM, LB.parameters, V.parameters, UNROLL)
     # return @show avx_body(ls, UNROLL)
     if last(UNROLL) > 1

src/simdfunctionals/filter.jl

@@ -8,8 +8,8 @@ function vfilter!(f::F, x::Vector{T}, y::AbstractArray{T}) where {F,T <: NativeT
     st = VectorizationBase.static_sizeof(T)
     zero_index = MM(W, Static(0), st)
     GC.@preserve x y begin
-        ptr_x = llvmptr(x)
-        ptr_y = llvmptr(y)
+        # ptr_x = llvmptr(x); ptr_y = llvmptr(y)
+        ptr_x = pointer(x); ptr_y = pointer(y)
         for _ ∈ 1:Nrep
             vy = VectorizationBase.__vload(ptr_y, zero_index, False(), register_size())
             mask = f(vy)