slowdown of atomics on 0.7 / fix llvmcall

[chethega]

Somehow all atomics became really slow and allocating. Example:

julia> using Base.Threads
julia> using BenchmarkTools
julia> a = Atomic{Int64}(0);

julia> @btime atomic_add!($a,$1);
  9.177 ns (0 allocations: 0 bytes)
julia> versioninfo()
Julia Version 0.6.3

julia> @btime atomic_add!($a,$1);
  129.263 ns (1 allocation: 16 bytes)
julia> versioninfo()
Julia Version 0.7.0-alpha.136

This is very visible when inspecting @code_llvm: Instead of the hand-written llvmcall, we get some weird giant blob of compiler output.

[vtjnash]

Looks like bad lowering (the call expression isn't getting linearized). The llvmcall "function" is one of the last special-case syntaxes that doesn't really behave properly. Similar to #26297

[JeffBezanson]

Adding to 1.0.x; a couple times I've run into cases where poking the compiler a bit causes it to blow up on llvmcall due to its un-settled IR form.

[chethega]

The following reproduces the bug and shows a workaround:

julia> using Base: llvmcall
julia> mutable struct foo
       val::Int64
       end
julia> function getval_1(x::foo)
       return llvmcall("%z = inttoptr i64 %0 to i64*\n %zz = load i64, i64* %z \n ret i64 %zz", Int64, Tuple{Ptr{Nothing}}, pointer_from_objref(x))
       end
julia> function getval_2(x::foo)
       ptr=pointer_from_objref(x); 
       return llvmcall("%z = inttoptr i64 %0 to i64*\n %zz = load i64, i64* %z \n ret i64 %zz", Int64, Tuple{Ptr{Nothing}}, ptr)
       end



julia> @code_llvm getval_2(a)
; Function getval_2
; Location: REPL[49]:2
define i64 @julia_getval_2_33505(%jl_value_t addrspace(10)* nonnull dereferenceable(8)) {
top:
; Function pointer_from_objref; {
; Location: pointer.jl:143
  %1 = addrspacecast %jl_value_t addrspace(10)* %0 to %jl_value_t addrspace(11)*
  %2 = addrspacecast %jl_value_t addrspace(11)* %1 to %jl_value_t*
;}
; Location: REPL[49]:3
  %z.i = bitcast %jl_value_t* %2 to i64*
  %zz.i = load i64, i64* %z.i, align 8
  ret i64 %zz.i
}


julia> @code_llvm getval_1(a)
#[truncated garbage]

julia> versioninfo()
Julia Version 0.7.0-beta.58
Commit 8e7e6fc00a (2018-06-27 18:34 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i5-5###U CPU @ 2.00GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.0 (ORCJIT, broadwell)

[mohamed82008]

Will/can this be addressed any time soon?

[mohamed82008]

Also

julia> VERSION
v"0.6.4"

julia> using BenchmarkTools

julia> l = Threads.SpinLock()
Base.Threads.TatasLock(Base.Threads.Atomic{Int64}(0))

julia> f(l) = (Threads.lock(l); Threads.unlock(l))
f (generic function with 1 method)

julia> @btime f($l)
  9.116 ns (0 allocations: 0 bytes)

julia> VERSION
v"0.7.0-beta.212"

julia> using BenchmarkTools

julia> l = Threads.SpinLock()
Base.Threads.TatasLock(Base.Threads.Atomic{Int64}(0))

julia> f(l) = (Threads.lock(l); Threads.unlock(l))
f (generic function with 1 method)

julia> @btime f($l)
  192.185 ns (3 allocations: 48 bytes)

[JeffBezanson]

Adding to 0.7 since this blocks SIMD.jl.