Ensure immutable isbits Union fields get set correctly #23367

quinnj · 2017-08-20T14:01:13Z

@vtjnash, this isn't quite right yet. When constructed, the selector byte and any non-union fields aren't getting set correctly.

I'm kind of just taking a stab here, and I'm guessing that my use of fval_info and dest are not quite right, hence the non-working solution here. For example, when I run

struct BitsUnion
    x::Union{Int64, Float64}
end
BitsUnion(5.0)

the fval_info is

(lldb) call jl_(fval_info.constant)
#<null>
(lldb) call jl_(fval_info.typ)
Float64
(lldb) p fval_info.isimmutable
(const bool) $0 = false

I'm not sure why the fval_info.constant isn't set, but that causes problems when we try to compute_tindex_unboxed later against BitsUnion.

I won't be able to do much else on this for the rest of the day, but any pointers/tips/direction you or @yuyichao could provide would be much appreciated.

fix #23351

quinnj · 2017-08-21T05:38:49Z

Ok, a little more sleuthing:

Turns out when you call julia_type_to_llvm on Union{Int64, Float64}, the resulting Type * has isPointerTy() == true
I believe that's why the value setting isn't happening correctly, since emit_unbox thinks it's dealing w/ a pointer type when it should be unboxing inline as a Float64 or Int64

I'm not quite sure what the right solution here is though. Modify emit_unbox? Lie to LLVM and call emit_unbox w/ the largest Union element?

I feel like there still might be another issue going on, because in the llvm code for BitsUnion(5.0), I still don't see a selector byte store for some reason. Would that be getting optimized out for some reason? Or "clobbered" somehow due to the Union-is-a-pointertype issue?

vtjnash · 2017-08-21T17:03:21Z

src/builtins.c

@@ -259,6 +259,10 @@ JL_CALLABLE(jl_f_sizeof)
 jl_value_t *x = args[0];
 if (jl_is_unionall(x) || jl_is_uniontype(x)) {
 x = jl_unwrap_unionall(x);
+ size_t elsize = 0, al = 0;
+ int isinline = jl_islayout_inline(x, &elsize, &al);


This will have the curious property that T == S will not imply sizeof(T) == sizeof(S). I'm not sure if that's going to be an issue, but it does make be nervous. In particular, consider Union{Tuple{Int64}, Tuple{Int8}}, which doesn't include selector bits, vs. Tuple{Union{Int64, Int8}}, which throws an error.

quinnj · 2017-08-21T23:04:43Z

Bump @vtjnash, I could really use your help/direction here.

quinnj · 2017-08-23T03:19:05Z

Alright, added some tests for @vtjnash's help w/ the fix. Let's see what CI says. Anyone else want to take a look?

quinnj · 2017-08-23T13:05:22Z

That's a lot of great green checkmarks; a shame that Travis is backed up so far.

vtjnash · 2017-08-23T16:46:49Z

Looks like Travis caught another bug though:

From worker 5:	compare_fields at /home/travis/build/JuliaLang/julia/src/builtins.c:94

(that line is assuming isbits_layout === isdatatype)

quinnj · 2017-08-24T13:37:43Z

Mac failure is #23215; everything else looks all good.

Keno · 2017-09-01T16:32:49Z

@vtjnash is this good to go?

JeffBezanson · 2017-09-01T16:35:34Z

src/rtutils.c

- depth);
+ jl_datatype_t *ft = (jl_datatype_t*)jl_field_type(vt, i);
+ if (jl_is_uniontype(ft)) {
+ uint8_t sel = ((uint8_t*)fld_ptr)[jl_field_size(vt, i) - 1];


This operation should be abstracted.

quinnj · 2017-09-02T04:20:08Z

I'm still seeing a segfault w/

struct A
    x::Union{Float64, Void}
end

getx(a::A) = a.x

getx(A(1.0))

quinnj · 2017-09-02T20:33:16Z

Ok, looks like the segfault is coming from

data = ctx.builder.CreateSelect(isboxed_union,
-> 5661	                                    ctx.builder.CreateLoad(retvalinfo.gcroot),
   5662	                                    data);

when retvalinfo.gcroot, which is NULL, is accessed. Note that I'm running master merged w/ this branch locally, in case that makes a difference.

I also see retvalinfo.ispointer() == true if that makes a difference.

(lldb) p retvalinfo.ispointer()
(bool) $11 = true
(lldb) call jl_(retvalinfo.typ)
Union{Float64, Int64}
(lldb) p retvalinfo.gcroot
(llvm::Value *) $12 = 0x0000000000000000
(lldb)

I'm not really sure what needs to be done here though.

Keno · 2017-09-03T20:43:35Z

@quinnj You might try rebasing on top of #23571, which cleans up some of this code. I'll take a look if it still crashes.

quinnj · 2017-09-04T12:29:27Z

That indeed seems to have fixed the problem @Keno. I've pushed this branch as rebased on top of #23571

quinnj · 2017-09-04T15:37:18Z

Hmmm, travis linux 64-bit & julia freebsd ci both stalled during tests; I imagine unrelated? Have we seen those kinds of hangs elsewhere?

nalimilan · 2017-09-04T16:42:56Z

Yes, I've seen a hang today in another PR.

When the new gc root placement pass was introduced, the .gcroot field lost its original meaning and became mostly superfluous. However, the codegen code for unions still used it to get a referenced to the boxed part of the split union. Clean that up by introducing an explicit field for that purpose and removing the .gcroot field.

anonymize all types, remove special representation for Ptr, fix Constant generation for types containing split-union fields

quinnj · 2017-09-05T02:03:03Z

Ok, rebased this again on top of @Keno 's new gcroot commit. Also added a test for the codgen-getfield crash I saw above. Let's see what CI says on all of this.

quinnj · 2017-09-05T13:11:54Z

CI all looks good, just the one hang on 32-bit travis linux. I'll let @Keno & @vtjnash decide when this is good to go. I'll keep doing more testing on my end.

Keno · 2017-09-05T17:12:30Z

src/llvm-late-gc-lowering.cpp

- CI->getType(), "", CI);
- ASCI->takeName(CI);
- CI->replaceAllUsesWith(ASCI);
+ auto *ptr = new PtrToIntInst(CI->getOperand(0), CI->getType(), "", CI);


This isn't valid by IR invariants (it'll complain on LLVM 5.0+). Why does this need to be an int?

That's what CI->getType() is. We can worry about LLVM 5 when we start supporting it.

@yuyichao is running in that configuration AFAIK.

I think 4.0 already has non-integral pointer so this will complain there already.

I'm on 4.0 but I do test svn from time to time to make sure it's ready whenever Arch's package is updated.

Should be enough to just keep the old addrspacecast and add an ptrtoint on top?

This comment is not addressed and it breaks LLVM 4.0 .

Keno · 2017-09-05T17:12:51Z

test/libgit2.jl

@@ -76,6 +76,8 @@ end
 a = Base.cconvert(Ptr{LibGit2.StrArrayStruct}, p)
 b = Base.unsafe_convert(Ptr{LibGit2.StrArrayStruct}, a)
 @test p == convert(Vector{String}, unsafe_load(b))
+ @noinline gcuse(a) = a


Should use Core.gcuse now.

StefanKarpinski · 2017-09-07T17:44:01Z

❤️

maleadt · 2017-09-07T22:45:57Z

This breaks llvmcall(fun::Ptr) pretty spectacularly:

$ gdb julia

(gdb) break ccall.cpp:1083
Breakpoint 1 (ccall.cpp:1083) pending.

(gdb) run
               _
   _       _ _(_)_     |  A fresh approach to technical computing
  (_)     | (_) (_)    |  Documentation: https://docs.julialang.org
   _ _   _| |_  __ _   |  Type "?help" for help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 0.7.0-DEV.1704 (2017-09-07 16:53 UTC)
 _/ |\__'_|_|_|\__'_|  |  Commit 1875957dec* (0 days old master)
|__/                   |  x86_64-unknown-linux-gnu

julia> x() = Base.llvmcall(C_NULL, Void, Tuple{Ptr{Int}, Int}, C_NULL, 1)
x (generic function with 1 method)

julia> x()

Thread 1 "julia" hit Breakpoint 1, emit_llvmcall (ctx=..., args=0x7fffeb1c6750, nargs=5) at /home/tbesard/Work/Julia-CUDA/julia-dev/src/ccall.cpp:1083
1083            assert(isPtr);

(gdb) call argtypes[0]->dump()
i64

(gdb) call argtypes[1]->dump()
i64

As seen on LLVM.jl, bisected to 1875957.

vtjnash · 2017-09-07T23:04:01Z

llvmcall is still experimental (unstable) anyways

maleadt · 2017-09-08T00:13:28Z

llvmcall is still experimental (unstable) anyways

Yeah sure stuff breaks on master all the time, just reporting the fact.

maleadt · 2017-09-08T17:27:51Z

OK, so looking at the actual commits this seems fully intended. Is this anonymization of pointers to native-width integers going to be the default from now on?
Even though llvmcall is unstable, this seems unnecessarily breaking..

ararslan requested review from vtjnash and yuyichao August 20, 2017 20:01

kshyatt added the domain:types and dispatch Types, subtyping and method dispatch label Aug 20, 2017

vtjnash reviewed Aug 21, 2017

View reviewed changes

vtjnash force-pushed the jq/bitsunions branch from fe66832 to 234e118 Compare August 22, 2017 17:52

quinnj force-pushed the jq/bitsunions branch from 8a66f03 to df0b9d0 Compare August 23, 2017 06:34

quinnj mentioned this pull request Aug 24, 2017

Preserve the input element type in unique #23208

Merged

vtjnash force-pushed the jq/bitsunions branch 6 times, most recently from 16863b9 to 7cf7580 Compare August 30, 2017 17:24

vtjnash changed the title ~~[WIP] Ensure immutable isbits Union fields get set correctly~~ Ensure immutable isbits Union fields get set correctly Aug 30, 2017

JeffBezanson reviewed Sep 1, 2017

View reviewed changes

quinnj force-pushed the jq/bitsunions branch from 7cf7580 to 4a5e85d Compare September 4, 2017 12:23

quinnj and others added 5 commits September 4, 2017 19:58

Work to ensure immutable isbits Union fields get set correctly. #23351

b83b7e9

correctly handle union-splitting in codegen type allocation

560aca9

Add tests for structs with isbits Union fields.

a562e33

correctly handle union-store-splitting in "is" and "object_id"

f61db81

fix codegen bugs, simplify type representations

2bb430e

anonymize all types, remove special representation for Ptr, fix Constant generation for types containing split-union fields

quinnj force-pushed the jq/bitsunions branch from 4a5e85d to 2bb430e Compare September 5, 2017 02:02

Keno reviewed Sep 5, 2017

View reviewed changes

vtjnash approved these changes Sep 5, 2017

View reviewed changes

quinnj merged commit 1875957 into master Sep 7, 2017

quinnj deleted the jq/bitsunions branch September 7, 2017 16:53

This was referenced Sep 8, 2017

Fix the generated example now that pointers are passes anonymously. maleadt/LLVM.jl#51

Merged

Adapt to new anonymous pointers in Julia. JuliaGPU/CUDAnative.jl#114

Merged

thomasfaingnaert mentioned this pull request Nov 14, 2019

Implement wrappers for WMMA LLVM intrinsics JuliaGPU/CUDAnative.jl#494

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Ensure immutable isbits Union fields get set correctly #23367

Ensure immutable isbits Union fields get set correctly #23367

quinnj commented Aug 20, 2017 •

edited by vtjnash

Loading

quinnj commented Aug 21, 2017

vtjnash Aug 21, 2017

quinnj commented Aug 21, 2017

quinnj commented Aug 23, 2017

quinnj commented Aug 23, 2017

vtjnash commented Aug 23, 2017

quinnj commented Aug 24, 2017

Keno commented Sep 1, 2017

JeffBezanson Sep 1, 2017

quinnj commented Sep 2, 2017

quinnj commented Sep 2, 2017 •

edited

Loading

Keno commented Sep 3, 2017

quinnj commented Sep 4, 2017

quinnj commented Sep 4, 2017

nalimilan commented Sep 4, 2017

quinnj commented Sep 5, 2017

quinnj commented Sep 5, 2017

Keno Sep 5, 2017

vtjnash Sep 5, 2017

Keno Sep 5, 2017

yuyichao Sep 5, 2017

yuyichao Sep 7, 2017

Keno Sep 5, 2017

StefanKarpinski commented Sep 7, 2017

maleadt commented Sep 7, 2017

vtjnash commented Sep 7, 2017

maleadt commented Sep 8, 2017

maleadt commented Sep 8, 2017 •

edited

Loading

Ensure immutable isbits Union fields get set correctly #23367

Ensure immutable isbits Union fields get set correctly #23367

Conversation

quinnj commented Aug 20, 2017 • edited by vtjnash Loading

quinnj commented Aug 21, 2017

Choose a reason for hiding this comment

quinnj commented Aug 21, 2017

quinnj commented Aug 23, 2017

quinnj commented Aug 23, 2017

vtjnash commented Aug 23, 2017

quinnj commented Aug 24, 2017

Keno commented Sep 1, 2017

Choose a reason for hiding this comment

quinnj commented Sep 2, 2017

quinnj commented Sep 2, 2017 • edited Loading

Keno commented Sep 3, 2017

quinnj commented Sep 4, 2017

quinnj commented Sep 4, 2017

nalimilan commented Sep 4, 2017

quinnj commented Sep 5, 2017

quinnj commented Sep 5, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

StefanKarpinski commented Sep 7, 2017

maleadt commented Sep 7, 2017

vtjnash commented Sep 7, 2017

maleadt commented Sep 8, 2017

maleadt commented Sep 8, 2017 • edited Loading

quinnj commented Aug 20, 2017 •

edited by vtjnash

Loading

quinnj commented Sep 2, 2017 •

edited

Loading

maleadt commented Sep 8, 2017 •

edited

Loading