Store to GC frame preventing vectorization

[yuyichao]

Using patched LLVM 3.7.1 and OrcJIT.

Bisect log

yuyichao% git bisect bad 
c3039d49bc53d17b31510f788eac6bc7c9ee7fff is the first bad commit
commit c3039d49bc53d17b31510f788eac6bc7c9ee7fff
Author: Jameson Nash <[email protected]>
Date:   Thu Mar 10 15:20:10 2016 -0500

    ensure that debug info is always available for -O0 mode

    this should ensure that the jlcall args array is always visible in the debugger
    and that all variables can be inspected from -O0 mode
    and that julia-debug is built at the -O0 optimization level (so all variables are visible)

:100644 100644 1ff6f833e6350d8a36372566c86a2124ff457d47 52cd40a56cd21b6b1472cd7aa6d5ca5f5741a338 M      Makefile
:040000 040000 7f2de288213c8055e7b122fb6d49889b5d8fd91f f0503a5bd854f1ae07c870991d8f7ba67e3c62e1 M      src

yuyichao% git bisect log 
git bisect start
# bad: [a5f2c7a7f2c0786a0fd4d0eebd5690bb905a349b] add a gc root and some write barriers needed from jb/linear3 merge
git bisect bad a5f2c7a7f2c0786a0fd4d0eebd5690bb905a349b
# good: [9bfd27bd380124174a5f37c342e5c048874d71a4] Merge pull request #13412 from JuliaLang/jb/functions
git bisect good 9bfd27bd380124174a5f37c342e5c048874d71a4
# good: [b252d6103bcd93878fa029674687bf4c94dd7f99] Merge pull request #15200 from JuliaLang/anj/cholp
git bisect good b252d6103bcd93878fa029674687bf4c94dd7f99
# good: [892b9406e3c68a56e43587de740520c16dd0f6d8] support multiple arguments in `Generator` by zipping
git bisect good 892b9406e3c68a56e43587de740520c16dd0f6d8
# bad: [a0135328997c0aaea0b0a8f38848aed903aa196f] Merge pull request #15567 from JuliaLang/yyc/llvm39
git bisect bad a0135328997c0aaea0b0a8f38848aed903aa196f
# bad: [f3374b9e381994dd845381cf5212779780d9377f] Merge pull request #15347 from JuliaLang/kf/refactorobjlookup
git bisect bad f3374b9e381994dd845381cf5212779780d9377f
# bad: [a75f4c31d808e10c8c4247d1f5797e17159450db] Merge pull request #15334 from omus/dateformat-docs
git bisect bad a75f4c31d808e10c8c4247d1f5797e17159450db
# good: [fc469b68316061d95003cb0768c43dbc4b1efd0f] Merge pull request #15462 from justbur/fix-15461
git bisect good fc469b68316061d95003cb0768c43dbc4b1efd0f
# bad: [1745a5fc4d4c377a52f19a752dbec306769408c8] Merge pull request #15443 from tlnagy/master
git bisect bad 1745a5fc4d4c377a52f19a752dbec306769408c8
# bad: [64dbecf6ca76a46be34c3addd39cd516f383b70a] Merge pull request #15444 from JuliaLang/jn/dwarfbug
git bisect bad 64dbecf6ca76a46be34c3addd39cd516f383b70a
# good: [ebf6c64a6ad566a42e4c8b5408f59e46b030e5f4] fix debug info in compiler
git bisect good ebf6c64a6ad566a42e4c8b5408f59e46b030e5f4
# bad: [c3039d49bc53d17b31510f788eac6bc7c9ee7fff] ensure that debug info is always available for -O0 mode
git bisect bad c3039d49bc53d17b31510f788eac6bc7c9ee7fff
# first bad commit: [c3039d49bc53d17b31510f788eac6bc7c9ee7fff] ensure that debug info is always available for -O0 mode

Possibly similar to #13301 but that was "fixed" after codegen_rewrite2 and SIMD doesn't even work for the cases that used to work before...

@vtjnash

[yuyichao]

Simple code and the corresponding IR on the first bad commit.

julia> function f(a)
           @inbounds @simd for i in eachindex(a)
               a[i] += 1
           end
           nothing
       end
f (generic function with 1 method)

full llvm-ir

if13:                                             ; preds = %if13.preheader, %if13
  %"##i#7433.01" = phi i64 [ %41, %if13 ], [ 0, %if13.preheader ]
  store %jl_value_t* %0, %jl_value_t** %3, align 8, !dbg !49, !llvm.mem.parallel_loop_access !57
  %35 = load float*, float** %24, align 8, !dbg !49, !tbaa !59, !llvm.mem.parallel_loop_access !57
  %36 = getelementptr float, float* %35, i64 %"##i#7433.01", !dbg !49
  %37 = load float, float* %36, align 4, !dbg !49, !tbaa !60, !llvm.mem.parallel_loop_access !57
  %38 = fadd float %37, 1.000000e+00, !dbg !49
  store %jl_value_t* %0, %jl_value_t** %4, align 8, !dbg !49, !llvm.mem.parallel_loop_access !57
  %39 = load float*, float** %24, align 8, !dbg !49, !tbaa !59, !llvm.mem.parallel_loop_access !57
  %40 = getelementptr float, float* %39, i64 %"##i#7433.01", !dbg !49
  store float %38, float* %40, align 4, !dbg !49, !tbaa !60, !llvm.mem.parallel_loop_access !57
  %41 = add nuw nsw i64 %"##i#7433.01", 1, !dbg !61, !simd_loop !13
  call void @llvm.dbg.value(metadata i64 %41, i64 0, metadata !22, metadata !31), !dbg !32
  %exitcond = icmp eq i64 %41, %22, !dbg !50
  br i1 %exitcond, label %L.backedge.loopexit, label %if13, !dbg !50, !llvm.loop !58

[KristofferC]

Ref #13777

[yuyichao]

Actually looks like the issue is similar to #15402 but it gets even worse after that

[yuyichao]

With #15735 and #13463 the example above vectorizes at normal optimization level. However, if the array is allocated in the same function or the variable is otherwise assgined to (similar to #13301) the redundant store in the loop still prevent the optimization from happening at normal optimization level.

function f_simd(n::Integer)
    a = zeros(Float32, n)
    @inbounds @simd for i in eachindex(a)
        a[i] += 1
    end
    nothing
end

IR of the inner loop:

if15:                                             ; preds = %if15, %if15.lr.ph
  %"i#256.017" = phi i64 [ 0, %if15.lr.ph ], [ %45, %if15 ]
  store %jl_value_t* %24, %jl_value_t** %9, align 8, !dbg !53, !tbaa !32, !llvm.mem.parallel_loop_access !59
  %42 = getelementptr float, float* %41, i64 %"i#256.017", !dbg !53
  %43 = load float, float* %42, align 4, !dbg !53, !tbaa !61, !llvm.mem.parallel_loop_access !59
  %44 = fadd float %43, 1.000000e+00, !dbg !53
  store %jl_value_t* %24, %jl_value_t** %10, align 8, !dbg !53, !tbaa !32, !llvm.mem.parallel_loop_access !59
  store float %44, float* %42, align 4, !dbg !53, !tbaa !61, !llvm.mem.parallel_loop_access !59
  %45 = add nuw nsw i64 %"i#256.017", 1, !dbg !62, !simd_loop !4
  call void @llvm.dbg.value(metadata i64 %45, i64 0, metadata !26, metadata !30), !dbg !31
  %exitcond = icmp eq i64 %45, %35, !dbg !58
  br i1 %exitcond, label %L11.loopexit, label %if15, !dbg !58, !llvm.loop !60

[yuyichao]

Vectorization works now (on llvm 3.8 at least) move to #15369