Why does Julia not optimize this code when C++ (LLVM) can?

Question

When using a C++ compiler with LLVM version 6.0.0, the following code

bool isEven(int n) {
    bool ret = true;
    for (int i = 0; i < n; i ++) {
        ret = !ret;
    }
    return ret;
}

emits the LLVM IR

define zeroext i1 @_Z6isEveni(i32) local_unnamed_addr #0 !dbg !7 {
  call void @llvm.dbg.value(metadata i32 %0, metadata !14, metadata !DIExpression()), !dbg !18
  call void @llvm.dbg.value(metadata i8 1, metadata !15, metadata !DIExpression()), !dbg !19
  call void @llvm.dbg.value(metadata i32 0, metadata !16, metadata !DIExpression()), !dbg !20
  %2 = icmp slt i32 %0, 1, !dbg !21
  %3 = and i32 %0, 1, !dbg !23
  %4 = icmp eq i32 %3, 0, !dbg !23
  %5 = or i1 %4, %2, !dbg !23
  ret i1 %5, !dbg !24
}

declare void @llvm.dbg.value(metadata, metadata, metadata) #1

attributes #0 = { nounwind readnone uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
attributes #1 = { nounwind readnone speculatable }

See: https://godbolt.org/z/oPBFey

This is functionally equivalent to the following implementation:

julia> isEven(n::Int) = rem(n, 2) != 0
isEven (generic function with 1 method)

julia> @code_llvm debuginfo=:none isEven(7)

define i8 @julia_isEven_18796(i64) {
top:
  %1 = trunc i64 %0 to i8
  %2 = and i8 %1, 1
  %3 = xor i8 %2, 1
  ret i8 %3
}

julia>

However, the original C++ implementation ported to Julia results in a very different LLVM IR:

julia> function isEven(n::Int)
           out = true
           for i in 0:n-1
               out = !out
           end
           return out
       end
isEven (generic function with 1 method)

julia> @code_llvm debuginfo=:none isEven(7)

define i8 @julia_isEven_18793(i64) {
top:
  %1 = add i64 %0, -1
  %2 = icmp sgt i64 %1, -1
  br i1 %2, label %L8.L12_crit_edge, label %L25

L8.L12_crit_edge:                                 ; preds = %top
  %min.iters.check = icmp ult i64 %0, 128
  br i1 %min.iters.check, label %scalar.ph, label %vector.ph

vector.ph:                                        ; preds = %L8.L12_crit_edge
  %n.vec = and i64 %0, -128
  br label %vector.body

vector.body:                                      ; preds = %vector.body, %vector.ph
  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
  %vec.phi = phi <32 x i8> [ <i8 1, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0>, %vector.ph ], [ %3, %vector.body ]
  %vec.phi8 = phi <32 x i8> [ zeroinitializer, %vector.ph ], [ %4, %vector.body ]
  %vec.phi9 = phi <32 x i8> [ zeroinitializer, %vector.ph ], [ %5, %vector.body ]
  %vec.phi10 = phi <32 x i8> [ zeroinitializer, %vector.ph ], [ %6, %vector.body ]
  %3 = xor <32 x i8> %vec.phi, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %4 = xor <32 x i8> %vec.phi8, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %5 = xor <32 x i8> %vec.phi9, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %6 = xor <32 x i8> %vec.phi10, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %index.next = add i64 %index, 128
  %7 = icmp eq i64 %index.next, %n.vec
  br i1 %7, label %middle.block, label %vector.body

middle.block:                                     ; preds = %vector.body
  %bin.rdx = xor <32 x i8> %vec.phi8, %vec.phi
  %bin.rdx14 = xor <32 x i8> %5, %bin.rdx
  %bin.rdx15 = xor <32 x i8> %6, %bin.rdx14
  %rdx.shuf = shufflevector <32 x i8> %bin.rdx15, <32 x i8> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %bin.rdx16 = xor <32 x i8> %bin.rdx15, %rdx.shuf
  %rdx.shuf17 = shufflevector <32 x i8> %bin.rdx16, <32 x i8> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %bin.rdx18 = xor <32 x i8> %bin.rdx16, %rdx.shuf17
  %rdx.shuf19 = shufflevector <32 x i8> %bin.rdx18, <32 x i8> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %bin.rdx20 = xor <32 x i8> %bin.rdx18, %rdx.shuf19
  %rdx.shuf21 = shufflevector <32 x i8> %bin.rdx20, <32 x i8> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %bin.rdx22 = xor <32 x i8> %bin.rdx20, %rdx.shuf21
  %rdx.shuf23 = shufflevector <32 x i8> %bin.rdx22, <32 x i8> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %bin.rdx24 = xor <32 x i8> %bin.rdx22, %rdx.shuf23
  %8 = extractelement <32 x i8> %bin.rdx24, i32 0
  %cmp.n = icmp eq i64 %n.vec, %0
  br i1 %cmp.n, label %L25, label %scalar.ph

scalar.ph:                                        ; preds = %middle.block, %L8.L12_crit_edge
  %bc.resume.val = phi i64 [ %n.vec, %middle.block ], [ 0, %L8.L12_crit_edge ]
  %bc.merge.rdx = phi i8 [ %8, %middle.block ], [ 1, %L8.L12_crit_edge ]
  br label %L12

L12:                                              ; preds = %L12, %scalar.ph
  %value_phi2 = phi i8 [ %bc.merge.rdx, %scalar.ph ], [ %9, %L12 ]
  %value_phi3 = phi i64 [ %bc.resume.val, %scalar.ph ], [ %11, %L12 ]
  %9 = xor i8 %value_phi2, 1
  %10 = icmp eq i64 %value_phi3, %1
  %11 = add i64 %value_phi3, 1
  br i1 %10, label %L25, label %L12

L25:                                              ; preds = %L12, %middle.block, %top
  %value_phi6 = phi i8 [ 1, %top ], [ %9, %L12 ], [ %8, %middle.block ]
  ret i8 %value_phi6
}


julia> versioninfo()
Julia Version 1.3.1
Commit 2d5741174c (2019-12-30 21:36 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin18.6.0)
  CPU: Intel(R) Core(TM) i7-7920HQ CPU @ 3.10GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, skylake)

julia>

Can anyone explain why Julia is not able to produce the same IR as a C++ compiler for essentially the same code with almost the same version of LLVM?

I think discourse.julialang.org might be a better place to ask this question. — carstenbauer, Feb 28 '20 at 23:37
So it seems like it's not a problem with the Julia code. I've posted it on discourse for discussion: https://discourse.julialang.org/t/why-does-julia-not-optimize-this-code-when-c-llvm-can/35314 — kdheepak, Feb 29 '20 at 03:22

score 2 · Answer 1 · answered Mar 04 '20 at 19:00

The short answer is: Julia and C++ are different language with different semantics and different compilers.

The different semantics means that different optimizations at legal. It would take some close looking to see if this was one that is legal in C++ but illegal in Julia. I would be surprised if it was.

The different compilers mean the compilers do different things. C++ compilers have had decades and probably hundreds of millions of dollars of developer time put into them (Even if a lot of it was donated by open source volunteers); even a younger compiler like Clang is still able to build directly on decades of proven ideas from older compilers like GCC.

The Julia compiler was first started in 2012. Much fewer hours have gone into it. Indeed I don't think it actually had its own optimizer at all until v0.6, which was 2017. LLVM does have an optimizer, that both Julia and Clang use. But they use it differently, they have different passes enabled and they provide it with different information (because of different semantics). Plus you are looking at the code before LLVM has run. (so might want to look at the assembly instread). The LLVM versions being the same between the two only matter for what instructions exist -- not for what optizations LLVM does since you are looking at the code before that gets to run.

Thanks for the answer! I was under the impression that the LLVM IR that was generated by Julia was after doing a "pass". But you are right, it doesn't do the optimization and I should really be looking at the machine code. — kdheepak, Mar 05 '20 at 01:19
I can't rule out that this LLVM IR has had 1 or more pass done over it. Would need to get someonewho have been deep into that part of the compiler to look at it. (and that is one of part I have never looked it. I think it will be in the codegen files. but not 100% sure.) — Frames Catherine White, Mar 05 '20 at 10:55

Why does Julia not optimize this code when C++ (LLVM) can?

1 Answers1