I'm pretty confused about what you're trying to accomplish beyond standard
closures. What is your ParameterHolder type for?
I rewrote your first example wrapping everything in a function. Is this
doing what you want it to? The LLVM looks fine.
function test(α)
k(u::Float64,t::Float64,α) = α*u
G = (u,t) -> k(u,t,1.01)
G2 = (u,t)->k(u,t,α)
const β = 1.01
G3 = (u,t)->k(u,t,β)
@code_llvm G(1., 2.)
@code_llvm G2(1., 2.)
@code_llvm G3(1., 2.)
end
julia> test(2.)
define double @"julia_#28_70100"(double, double) #0 {
top:
%2 = fmul double %0, 1.010000e+00
ret double %2
}
define double @"julia_#29_70102"(%"##29#33"*, double, double) #0 {
top:
%3 = getelementptr inbounds %"##29#33", %"##29#33"* %0, i64 0, i32 0
%4 = load double, double* %3, align 8
%5 = fmul double %4, %1
ret double %5
}
define double @"julia_#30_70104"(%"##30#34"*, double, double) #0 {
top:
%3 = getelementptr inbounds %"##30#34", %"##30#34"* %0, i64 0, i32 1
%4 = load double, double* %3, align 8
%5 = fmul double %4, %1
ret double %5
}
On Tuesday, August 23, 2016 at 6:00:37 PM UTC-4, Chris Rackauckas wrote:
>
> Yes, I am looking for a closure which has the least overhead possible
> (this was all in v0.5). For example, for re-ordering parameters: g =
> (du,u,t) -> f(t,u,du), or for enclosing parameter values as above.
>
> I'll give the Val method a try and see whether the compile time is
> significant (it will probably be an option).
>
> Even a solution which is like translator1/2 where the inbounds check is
> able to be turned off would likely be performant enough to not be
> noticeably different.
>
> On Tuesday, August 23, 2016 at 1:27:25 PM UTC-7, Erik Schnetter wrote:
>>
>> Chris
>>
>> I don't quite understand what you mean. Are you looking for a closure /
>> lambda expression?
>>
>> ```Julia
>> function myfunc(x0, x1, alpha)
>> f(x) = alpha * x
>> ODE.solve(f, x0, x1)
>> end
>> ```
>>
>> Or is it important for your that your function `f` is optimized, i.e. you
>> want to re-run the code generator (expensive!) every time there's a new
>> value for `alpha`? For this, you can use `Val` (but please benchmark first):
>>
>> ```Julia
>> function f{alpha}(x, ::Type{Val{alpha}})
>> alpha * x
>> end
>>
>> function myfunc(x0, x1, alpha)
>> f1(x) = f(x, Val{alpha})
>> ODE.solve(f, x0, x1)
>> end
>> ```
>>
>> This will have a marginally faster evaluation of `f`, at the cost of
>> compiling a separate function for each value of `alpha`.
>>
>> Since these examples use closures, they will be much more efficient in
>> Julia 0.5 than in 0.4.
>>
>> -erik
>>
>>
>>
>>
>> On Tue, Aug 23, 2016 at 2:23 PM, Chris Rackauckas <[email protected]>
>> wrote:
>>
>>> Note: This looks long, but really just has a lot of LLVM IR!
>>>
>>> I have been digging into the issue recently of the best way to enclose
>>> parameters with a function
>>> <https://github.com/ChrisRackauckas/DifferentialEquations.jl/issues/41>.
>>> This is an issue that comes up a lot with scientific codes, and so I was
>>> hoping to try and get it right. However, the results of my experiments
>>> aren't looking too good, and so I was hoping to find out whether I am
>>> running into some bug or simply just not finding the optimal solution.
>>>
>>> The example is as follows (with LLVM IR included to show how exactly
>>> everything is compiling). Say the user wants we to do a bunch of things
>>> with the function f(u,t)=α*u where α is some parameter. They don't
>>> necessarily want to replace it as a constant since they may change it
>>> around a bit, but every time this function is given to me, I can treat it
>>> as a constant. If they were willing to treat it as a constant, then they
>>> could take this function:
>>>
>>> k(u::Float64,t::Float64,α) = α*u
>>> println("Standard k definition")
>>> @code_llvm k(1.0,2.0,1.01)
>>>
>>> #Result
>>>
>>> define double @julia_k_70163(double, double, double) #0 {
>>> top:
>>> %3 = fmul double %0, %2
>>> ret double %3
>>> }
>>>
>>>
>>> and enclose the constant:
>>>
>>> G = (u,t) -> k(u,t,1.01)
>>> G2 = (u,t)->k(u,t,α)
>>> println("Top level inlined k")
>>> @code_llvm G(1.0,2.0)
>>> println("Top level not inlined k")
>>> @code_llvm G2(1.0,2.0)
>>> const β = 1.01
>>> G3 = (u,t)->k(u,t,β)
>>> println("Top level not inlined but const k")
>>> @code_llvm G3(1.0,2.0)
>>>
>>> #Results
>>>
>>> Top level inlined k
>>>
>>> define double @"julia_#159_70165"(double, double) #0 {
>>> top:
>>> %2 = fmul double %0, 1.010000e+00
>>> ret double %2
>>> }
>>>
>>> Top level not inlined k
>>>
>>> define %jl_value_t* @"julia_#161_70167"(double, double) #0 {
>>> top:
>>> %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #2
>>> %ptls_i8 = getelementptr i8, i8* %thread_ptr, i64 -2672
>>> %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
>>> %2 = alloca [5 x %jl_value_t*], align 8
>>> %.sub = getelementptr inbounds [5 x %jl_value_t*], [5 x %jl_value_t*]*
>>> %2, i64 0, i64 0
>>> %3 = getelementptr [5 x %jl_value_t*], [5 x %jl_value_t*]* %2, i64 0,
>>> i64 2
>>> %4 = bitcast %jl_value_t** %3 to i8*
>>> call void @llvm.memset.p0i8.i32(i8* %4, i8 0, i32 24, i32 8, i1 false)
>>> %5 = bitcast [5 x %jl_value_t*]* %2 to i64*
>>> store i64 6, i64* %5, align 8
>>> %6 = getelementptr [5 x %jl_value_t*], [5 x %jl_value_t*]* %2, i64 0,
>>> i64 1
>>> %7 = bitcast i8* %ptls_i8 to i64*
>>> %8 = load i64, i64* %7, align 8
>>> %9 = bitcast %jl_value_t** %6 to i64*
>>> store i64 %8, i64* %9, align 8
>>> store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
>>> %10 = getelementptr [5 x %jl_value_t*], [5 x %jl_value_t*]* %2, i64 0,
>>> i64 4
>>> %11 = getelementptr [5 x %jl_value_t*], [5 x %jl_value_t*]* %2, i64 0,
>>> i64 3
>>> %12 = load i64, i64* inttoptr (i64 139896404414328 to i64*), align 8
>>> %13 = bitcast %jl_value_t** %11 to i64*
>>> store i64 %12, i64* %13, align 8
>>> store %jl_value_t* inttoptr (i64 139896327403528 to %jl_value_t*),
>>> %jl_value_t** %3, align 8
>>> %14 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %15 = getelementptr inbounds %jl_value_t, %jl_value_t* %14, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %15, align 8
>>> %16 = bitcast %jl_value_t* %14 to double*
>>> store double %0, double* %16, align 8
>>> store %jl_value_t* %14, %jl_value_t** %10, align 8
>>> %17 = call %jl_value_t* @jl_apply_generic(%jl_value_t** %3, i32 3)
>>> %18 = load i64, i64* %9, align 8
>>> store i64 %18, i64* %7, align 8
>>> ret %jl_value_t* %17
>>> }
>>>
>>> Top level not inlined but const k
>>>
>>> define double @"julia_#163_70169"(double, double) #0 {
>>> top:
>>> %2 = fmul double %0, 1.010000e+00
>>> ret double %2
>>> }
>>>
>>> It's clear from these results that the user would have to treat that
>>> values they enclose as a constant. This is a known performance issue due to
>>> global variables.
>>>
>>> So let's say I will be doing the enclosing on my end. The parameters
>>> ends up inside my function, and I am willing to wrap that into another
>>> function which holds all the parameters (can be necessary for calling some
>>> C libraries). However, I ran into issues finding any option which was
>>> performant. Here's my attempts:
>>>
>>> # Some helpers
>>> immutable ParameterHolder{uType<:Number}
>>> α::uType
>>> end
>>>
>>> function translator(f,α)
>>> Base.@propagate_inbounds g(u::Float64,t) = f(u::Float64,t,α)
>>> return g
>>> end
>>>
>>> function translator2(f,α)
>>> @inbounds g = (u::Float64,t) -> f(u::Float64,t,α)
>>> return g
>>> end
>>>
>>>
>>> # The "main" function the user will call
>>> function code_test()
>>> local u::Float64 = 1.0
>>> const pconst = ParameterHolder(1.01)
>>> p = ParameterHolder(1.01)
>>>
>>> f(u,t,p) = @inbounds return p.α*u
>>> println("Inside using ParameterHolder Constant")
>>> @code_llvm f(1.0,2.0,pconst)
>>> println("Inside using ParameterHolder")
>>> @code_llvm f(1.0,2.0,p)
>>> h(u,t,α) = α*u
>>> α = 1.01
>>> g = (u,t) -> h(u,t,α)
>>> println("Inside using closure with variable")
>>> @code_llvm g(u,2.0)
>>> const β = 1.01
>>> l = (u,t) -> h(u,t,β)
>>> println("Inside using closure with const variable")
>>> @code_llvm l(u,2.0)
>>> m = (u,t) -> k(u,t,β)
>>> println("Inside using outside function closure with variable")
>>> @code_llvm m(u,2.0)
>>> J = (u::Float64,t::Float64) -> k(u::Float64,t::Float64,α::Float64)
>>> println("Inside using outside function closure with variable")
>>> @code_llvm J(u,2.0)
>>> @inline J2(u::Float64,t::Float64) =
>>> k(u::Float64,t::Float64,1.01::Float64)
>>> println("Inside using closure inlined")
>>> @code_llvm J2(u::Float64,2.0)
>>> J3 = translator(k,1.01)
>>> println("Inside using translator")
>>> @code_llvm J3(u::Float64,2.0)
>>> println(J3(u,2.0))
>>> J4 = translator(g,1.01)
>>> println("Inside using translator of inside")
>>> @code_llvm J4(u::Float64,2.0)
>>> J5 = translator(k,1.01)
>>> println("Inside using translator 2")
>>> @code_llvm J5(u::Float64,2.0)
>>> println(J3(u,2.0))
>>> J6 = translator(g,1.01)
>>> println("Inside using translator 2 of inside")
>>> @code_llvm J6(u::Float64,2.0)
>>> end
>>> code_test()
>>>
>>> Let's walk through all of the results. If the user gave me the
>>> parameters in an immutable ParameterHolder, I can get functions like:
>>>
>>> Inside using ParameterHolder Constant
>>>
>>> define double @julia_f_70172(double, double, %ParameterHolder*) #0 {
>>> top:
>>> %3 = getelementptr inbounds %ParameterHolder, %ParameterHolder* %2,
>>> i64 0, i32 0
>>> %4 = load double, double* %3, align 8
>>> %5 = fmul double %4, %0
>>> ret double %5
>>> }
>>> Inside using ParameterHolder
>>>
>>> define double @julia_f_70172(double, double, %ParameterHolder*) #0 {
>>> top:
>>> %3 = getelementptr inbounds %ParameterHolder, %ParameterHolder* %2,
>>> i64 0, i32 0
>>> %4 = load double, double* %3, align 8
>>> %5 = fmul double %4, %0
>>> ret double %5
>>> }
>>>
>>> This has more steps than necessary, but is okay. The ways using a
>>> closure fair rather poorly. For some reason, even though u is type-stable,
>>> it compiles functions for jl_value_t's:
>>>
>>> Inside using closure
>>>
>>> define %jl_value_t* @"julia_#146_70175"(%jl_value_t*, double, double) #0
>>> {
>>> top:
>>> %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #2
>>> %ptls_i8 = getelementptr i8, i8* %thread_ptr, i64 -2672
>>> %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
>>> %3 = alloca [10 x %jl_value_t*], align 8
>>> %.sub = getelementptr inbounds [10 x %jl_value_t*], [10 x
>>> %jl_value_t*]* %3, i64 0, i64 0
>>> %4 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 2
>>> %5 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 8
>>> %6 = bitcast %jl_value_t** %4 to i8*
>>> call void @llvm.memset.p0i8.i32(i8* %6, i8 0, i32 64, i32 8, i1 false)
>>> %7 = bitcast [10 x %jl_value_t*]* %3 to i64*
>>> store i64 16, i64* %7, align 8
>>> %8 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 1
>>> %9 = bitcast i8* %ptls_i8 to i64*
>>> %10 = load i64, i64* %9, align 8
>>> %11 = bitcast %jl_value_t** %8 to i64*
>>> store i64 %10, i64* %11, align 8
>>> store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
>>> %12 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 6
>>> %13 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 5
>>> %14 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 4
>>> %15 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 3
>>> %16 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 7
>>> %17 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 9
>>> %18 = getelementptr inbounds %jl_value_t, %jl_value_t* %0, i64 1, i32 0
>>> %19 = bitcast %jl_value_t** %18 to i64*
>>> %20 = load i64, i64* %19, align 8
>>> %21 = bitcast %jl_value_t** %5 to i64*
>>> store i64 %20, i64* %21, align 8
>>> store %jl_value_t* inttoptr (i64 139896320318056 to %jl_value_t*),
>>> %jl_value_t** %17, align 8
>>> %22 = call %jl_value_t* @jl_f_getfield(%jl_value_t* null,
>>> %jl_value_t** %5, i32 2)
>>> store %jl_value_t* %22, %jl_value_t** %4, align 8
>>> %23 = bitcast %jl_value_t* %0 to i64*
>>> %24 = load i64, i64* %23, align 8
>>> %25 = bitcast %jl_value_t** %12 to i64*
>>> store i64 %24, i64* %25, align 8
>>> store %jl_value_t* inttoptr (i64 139896320318056 to %jl_value_t*),
>>> %jl_value_t** %16, align 8
>>> %26 = call %jl_value_t* @jl_f_getfield(%jl_value_t* null,
>>> %jl_value_t** %12, i32 2)
>>> store %jl_value_t* %26, %jl_value_t** %13, align 8
>>> %27 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %28 = getelementptr inbounds %jl_value_t, %jl_value_t* %27, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %28, align 8
>>> %29 = bitcast %jl_value_t* %27 to double*
>>> store double %1, double* %29, align 8
>>> store %jl_value_t* %27, %jl_value_t** %15, align 8
>>> %30 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %31 = getelementptr inbounds %jl_value_t, %jl_value_t* %30, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %31, align 8
>>> %32 = bitcast %jl_value_t* %30 to double*
>>> store double %2, double* %32, align 8
>>> store %jl_value_t* %30, %jl_value_t** %14, align 8
>>> %33 = call %jl_value_t* @jl_apply_generic(%jl_value_t** %4, i32 4)
>>> %34 = load i64, i64* %11, align 8
>>> store i64 %34, i64* %9, align 8
>>> ret %jl_value_t* %33
>>> }
>>> Inside using closure with variable
>>>
>>> define %jl_value_t* @"julia_#147_70177"(%jl_value_t*, double, double) #0
>>> {
>>> top:
>>> %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #2
>>> %ptls_i8 = getelementptr i8, i8* %thread_ptr, i64 -2672
>>> %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
>>> %3 = alloca [10 x %jl_value_t*], align 8
>>> %.sub = getelementptr inbounds [10 x %jl_value_t*], [10 x
>>> %jl_value_t*]* %3, i64 0, i64 0
>>> %4 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 2
>>> %5 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 8
>>> %6 = bitcast %jl_value_t** %4 to i8*
>>> call void @llvm.memset.p0i8.i32(i8* %6, i8 0, i32 64, i32 8, i1 false)
>>> %7 = bitcast [10 x %jl_value_t*]* %3 to i64*
>>> store i64 16, i64* %7, align 8
>>> %8 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 1
>>> %9 = bitcast i8* %ptls_i8 to i64*
>>> %10 = load i64, i64* %9, align 8
>>> %11 = bitcast %jl_value_t** %8 to i64*
>>> store i64 %10, i64* %11, align 8
>>> store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
>>> %12 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 6
>>> %13 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 5
>>> %14 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 4
>>> %15 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 3
>>> %16 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 7
>>> %17 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 9
>>> %18 = bitcast %jl_value_t* %0 to i64*
>>> %19 = load i64, i64* %18, align 8
>>> %20 = bitcast %jl_value_t** %5 to i64*
>>> store i64 %19, i64* %20, align 8
>>> store %jl_value_t* inttoptr (i64 139896320318056 to %jl_value_t*),
>>> %jl_value_t** %17, align 8
>>> %21 = call %jl_value_t* @jl_f_getfield(%jl_value_t* null,
>>> %jl_value_t** %5, i32 2)
>>> store %jl_value_t* %21, %jl_value_t** %4, align 8
>>> %22 = getelementptr inbounds %jl_value_t, %jl_value_t* %0, i64 1, i32 0
>>> %23 = bitcast %jl_value_t** %22 to i64*
>>> %24 = load i64, i64* %23, align 8
>>> %25 = bitcast %jl_value_t** %12 to i64*
>>> store i64 %24, i64* %25, align 8
>>> store %jl_value_t* inttoptr (i64 139896320318056 to %jl_value_t*),
>>> %jl_value_t** %16, align 8
>>> %26 = call %jl_value_t* @jl_f_getfield(%jl_value_t* null,
>>> %jl_value_t** %12, i32 2)
>>> store %jl_value_t* %26, %jl_value_t** %13, align 8
>>> %27 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %28 = getelementptr inbounds %jl_value_t, %jl_value_t* %27, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %28, align 8
>>> %29 = bitcast %jl_value_t* %27 to double*
>>> store double %1, double* %29, align 8
>>> store %jl_value_t* %27, %jl_value_t** %15, align 8
>>> %30 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %31 = getelementptr inbounds %jl_value_t, %jl_value_t* %30, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %31, align 8
>>> %32 = bitcast %jl_value_t* %30 to double*
>>> store double %2, double* %32, align 8
>>> store %jl_value_t* %30, %jl_value_t** %14, align 8
>>> %33 = call %jl_value_t* @jl_apply_generic(%jl_value_t** %4, i32 4)
>>> %34 = load i64, i64* %11, align 8
>>> store i64 %34, i64* %9, align 8
>>> ret %jl_value_t* %33
>>> }
>>> Inside using closure with const variable
>>>
>>> define %jl_value_t* @"julia_#148_70179"(%jl_value_t*, double, double) #0
>>> {
>>> top:
>>> %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #2
>>> %ptls_i8 = getelementptr i8, i8* %thread_ptr, i64 -2672
>>> %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
>>> %3 = alloca [10 x %jl_value_t*], align 8
>>> %.sub = getelementptr inbounds [10 x %jl_value_t*], [10 x
>>> %jl_value_t*]* %3, i64 0, i64 0
>>> %4 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 2
>>> %5 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 8
>>> %6 = bitcast %jl_value_t** %4 to i8*
>>> call void @llvm.memset.p0i8.i32(i8* %6, i8 0, i32 64, i32 8, i1 false)
>>> %7 = bitcast [10 x %jl_value_t*]* %3 to i64*
>>> store i64 16, i64* %7, align 8
>>> %8 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 1
>>> %9 = bitcast i8* %ptls_i8 to i64*
>>> %10 = load i64, i64* %9, align 8
>>> %11 = bitcast %jl_value_t** %8 to i64*
>>> store i64 %10, i64* %11, align 8
>>> store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
>>> %12 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 6
>>> %13 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 5
>>> %14 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 4
>>> %15 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 3
>>> %16 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 7
>>> %17 = getelementptr [10 x %jl_value_t*], [10 x %jl_value_t*]* %3, i64
>>> 0, i64 9
>>> %18 = getelementptr inbounds %jl_value_t, %jl_value_t* %0, i64 1, i32 0
>>> %19 = bitcast %jl_value_t** %18 to i64*
>>> %20 = load i64, i64* %19, align 8
>>> %21 = bitcast %jl_value_t** %5 to i64*
>>> store i64 %20, i64* %21, align 8
>>> store %jl_value_t* inttoptr (i64 139896320318056 to %jl_value_t*),
>>> %jl_value_t** %17, align 8
>>> %22 = call %jl_value_t* @jl_f_getfield(%jl_value_t* null,
>>> %jl_value_t** %5, i32 2)
>>> store %jl_value_t* %22, %jl_value_t** %4, align 8
>>> %23 = bitcast %jl_value_t* %0 to i64*
>>> %24 = load i64, i64* %23, align 8
>>> %25 = bitcast %jl_value_t** %12 to i64*
>>> store i64 %24, i64* %25, align 8
>>> store %jl_value_t* inttoptr (i64 139896320318056 to %jl_value_t*),
>>> %jl_value_t** %16, align 8
>>> %26 = call %jl_value_t* @jl_f_getfield(%jl_value_t* null,
>>> %jl_value_t** %12, i32 2)
>>> store %jl_value_t* %26, %jl_value_t** %13, align 8
>>> %27 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %28 = getelementptr inbounds %jl_value_t, %jl_value_t* %27, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %28, align 8
>>> %29 = bitcast %jl_value_t* %27 to double*
>>> store double %1, double* %29, align 8
>>> store %jl_value_t* %27, %jl_value_t** %15, align 8
>>> %30 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %31 = getelementptr inbounds %jl_value_t, %jl_value_t* %30, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %31, align 8
>>> %32 = bitcast %jl_value_t* %30 to double*
>>> store double %2, double* %32, align 8
>>> store %jl_value_t* %30, %jl_value_t** %14, align 8
>>> %33 = call %jl_value_t* @jl_apply_generic(%jl_value_t** %4, i32 4)
>>> %34 = load i64, i64* %11, align 8
>>> store i64 %34, i64* %9, align 8
>>> ret %jl_value_t* %33
>>> }
>>> Inside using outside function closure with variable
>>>
>>> define %jl_value_t* @"julia_#149_70181"(%jl_value_t*, double, double) #0
>>> {
>>> top:
>>> %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #2
>>> %ptls_i8 = getelementptr i8, i8* %thread_ptr, i64 -2672
>>> %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
>>> %3 = alloca [7 x %jl_value_t*], align 8
>>> %.sub = getelementptr inbounds [7 x %jl_value_t*], [7 x %jl_value_t*]*
>>> %3, i64 0, i64 0
>>> %4 = getelementptr [7 x %jl_value_t*], [7 x %jl_value_t*]* %3, i64 0,
>>> i64 2
>>> %5 = getelementptr [7 x %jl_value_t*], [7 x %jl_value_t*]* %3, i64 0,
>>> i64 5
>>> %6 = bitcast %jl_value_t** %4 to i8*
>>> call void @llvm.memset.p0i8.i32(i8* %6, i8 0, i32 40, i32 8, i1 false)
>>> %7 = bitcast [7 x %jl_value_t*]* %3 to i64*
>>> store i64 10, i64* %7, align 8
>>> %8 = getelementptr [7 x %jl_value_t*], [7 x %jl_value_t*]* %3, i64 0,
>>> i64 1
>>> %9 = bitcast i8* %ptls_i8 to i64*
>>> %10 = load i64, i64* %9, align 8
>>> %11 = bitcast %jl_value_t** %8 to i64*
>>> store i64 %10, i64* %11, align 8
>>> store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
>>> %12 = getelementptr [7 x %jl_value_t*], [7 x %jl_value_t*]* %3, i64 0,
>>> i64 4
>>> %13 = getelementptr [7 x %jl_value_t*], [7 x %jl_value_t*]* %3, i64 0,
>>> i64 3
>>> %14 = getelementptr [7 x %jl_value_t*], [7 x %jl_value_t*]* %3, i64 0,
>>> i64 6
>>> %15 = bitcast %jl_value_t* %0 to i64*
>>> %16 = load i64, i64* %15, align 8
>>> %17 = bitcast %jl_value_t** %5 to i64*
>>> store i64 %16, i64* %17, align 8
>>> store %jl_value_t* inttoptr (i64 139896320318056 to %jl_value_t*),
>>> %jl_value_t** %14, align 8
>>> %18 = call %jl_value_t* @jl_f_getfield(%jl_value_t* null,
>>> %jl_value_t** %5, i32 2)
>>> store %jl_value_t* %18, %jl_value_t** %13, align 8
>>> store %jl_value_t* inttoptr (i64 139896327403528 to %jl_value_t*),
>>> %jl_value_t** %4, align 8
>>> %19 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %20 = getelementptr inbounds %jl_value_t, %jl_value_t* %19, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %20, align 8
>>> %21 = bitcast %jl_value_t* %19 to double*
>>> store double %1, double* %21, align 8
>>> store %jl_value_t* %19, %jl_value_t** %12, align 8
>>> %22 = call %jl_value_t* @jl_apply_generic(%jl_value_t** %4, i32 3)
>>> %23 = load i64, i64* %11, align 8
>>> store i64 %23, i64* %9, align 8
>>> ret %jl_value_t* %22
>>> }
>>> Inside using outside function closure with variable
>>>
>>> define double @"julia_#150_70183"(%jl_value_t*, double, double) #0 {
>>> top:
>>> %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #3
>>> %ptls_i8 = getelementptr i8, i8* %thread_ptr, i64 -2672
>>> %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
>>> %3 = alloca [5 x %jl_value_t*], align 8
>>> %.sub = getelementptr inbounds [5 x %jl_value_t*], [5 x %jl_value_t*]*
>>> %3, i64 0, i64 0
>>> %4 = getelementptr [5 x %jl_value_t*], [5 x %jl_value_t*]* %3, i64 0,
>>> i64 3
>>> %5 = getelementptr [5 x %jl_value_t*], [5 x %jl_value_t*]* %3, i64 0,
>>> i64 2
>>> %6 = bitcast %jl_value_t** %4 to i8*
>>> call void @llvm.memset.p0i8.i32(i8* %6, i8 0, i32 16, i32 8, i1 false)
>>> %7 = bitcast [5 x %jl_value_t*]* %3 to i64*
>>> store i64 6, i64* %7, align 8
>>> %8 = getelementptr [5 x %jl_value_t*], [5 x %jl_value_t*]* %3, i64 0,
>>> i64 1
>>> %9 = bitcast i8* %ptls_i8 to i64*
>>> %10 = load i64, i64* %9, align 8
>>> %11 = bitcast %jl_value_t** %8 to i64*
>>> store i64 %10, i64* %11, align 8
>>> store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
>>> store %jl_value_t* null, %jl_value_t** %5, align 8
>>> %12 = getelementptr [5 x %jl_value_t*], [5 x %jl_value_t*]* %3, i64 0,
>>> i64 4
>>> %13 = bitcast %jl_value_t* %0 to i64*
>>> %14 = load i64, i64* %13, align 8
>>> %15 = bitcast %jl_value_t** %4 to i64*
>>> store i64 %14, i64* %15, align 8
>>> store %jl_value_t* inttoptr (i64 139896320318056 to %jl_value_t*),
>>> %jl_value_t** %12, align 8
>>> %16 = call %jl_value_t* @jl_f_getfield(%jl_value_t* null,
>>> %jl_value_t** %4, i32 2)
>>> store %jl_value_t* %16, %jl_value_t** %5, align 8
>>> %17 = getelementptr inbounds %jl_value_t, %jl_value_t* %16, i64 -1,
>>> i32 0
>>> %18 = bitcast %jl_value_t** %17 to i64*
>>> %19 = load i64, i64* %18, align 8
>>> %20 = and i64 %19, -16
>>> %21 = inttoptr i64 %20 to %jl_value_t*
>>> %22 = icmp eq %jl_value_t* %21, inttoptr (i64 139896322417392 to
>>> %jl_value_t*)
>>> br i1 %22, label %pass, label %fail
>>>
>>> fail: ; preds = %top
>>> call void @jl_type_error_rt(i8* inttoptr (i64 116848560 to i8*), i8*
>>> inttoptr (i64 64818736 to i8*), %jl_value_t* inttoptr (i64 139896322417392
>>> to %jl_value_t*), %jl_value_t* %16)
>>> unreachable
>>>
>>> pass: ; preds = %top
>>> %23 = bitcast %jl_value_t* %16 to double*
>>> %24 = load double, double* %23, align 16
>>> %25 = fmul double %24, %1
>>> %26 = load i64, i64* %11, align 8
>>> store i64 %26, i64* %9, align 8
>>> ret double %25
>>> }
>>>
>>>
>>> The only way to fix this is to manually inline the number as in J2:
>>>
>>> Inside using closure inlined
>>>
>>> define double @julia_J2_70185(double, double) #0 {
>>> top:
>>> %2 = fmul double %0, 1.010000e+00
>>> ret double %2
>>> }
>>>
>>> Note that even @inline failed to generate suitable code. What's
>>> interesting is that using the translator function tended to work okay. But
>>> the results show that this trick is only good for externally defined
>>> functions:
>>>
>>> Inside using translator
>>>
>>> define double @julia_g_70187(%"#g#143"*, double, double) #0 {
>>> top:
>>> %3 = getelementptr inbounds %"#g#143", %"#g#143"* %0, i64 0, i32 1
>>> %4 = load double, double* %3, align 8
>>> %5 = fmul double %4, %1
>>> ret double %5
>>> }
>>> 1.01
>>> Inside using translator of inside
>>>
>>> define %jl_value_t* @julia_g_70316(%jl_value_t*, double, double) #0 {
>>> top:
>>> %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #2
>>> %ptls_i8 = getelementptr i8, i8* %thread_ptr, i64 -2672
>>> %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
>>> %3 = alloca [6 x %jl_value_t*], align 8
>>> %.sub = getelementptr inbounds [6 x %jl_value_t*], [6 x %jl_value_t*]*
>>> %3, i64 0, i64 0
>>> %4 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 2
>>> %5 = bitcast %jl_value_t** %4 to i8*
>>> call void @llvm.memset.p0i8.i32(i8* %5, i8 0, i32 32, i32 8, i1 false)
>>> %6 = bitcast [6 x %jl_value_t*]* %3 to i64*
>>> store i64 8, i64* %6, align 8
>>> %7 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 1
>>> %8 = bitcast i8* %ptls_i8 to i64*
>>> %9 = load i64, i64* %8, align 8
>>> %10 = bitcast %jl_value_t** %7 to i64*
>>> store i64 %9, i64* %10, align 8
>>> store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
>>> %11 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 5
>>> %12 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 4
>>> %13 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 3
>>> %14 = bitcast %jl_value_t* %0 to i64*
>>> %15 = load i64, i64* %14, align 8
>>> %16 = getelementptr %jl_value_t, %jl_value_t* %0, i64 1
>>> %17 = bitcast %jl_value_t* %16 to i64*
>>> %18 = load i64, i64* %17, align 8
>>> %19 = bitcast %jl_value_t** %4 to i64*
>>> store i64 %15, i64* %19, align 8
>>> %20 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %21 = getelementptr inbounds %jl_value_t, %jl_value_t* %20, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %21, align 8
>>> %22 = bitcast %jl_value_t* %20 to double*
>>> store double %1, double* %22, align 8
>>> store %jl_value_t* %20, %jl_value_t** %13, align 8
>>> %23 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %24 = getelementptr inbounds %jl_value_t, %jl_value_t* %23, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %24, align 8
>>> %25 = bitcast %jl_value_t* %23 to double*
>>> store double %2, double* %25, align 8
>>> store %jl_value_t* %23, %jl_value_t** %12, align 8
>>> %26 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %27 = getelementptr inbounds %jl_value_t, %jl_value_t* %26, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %27, align 8
>>> %28 = bitcast %jl_value_t* %26 to i64*
>>> store i64 %18, i64* %28, align 8
>>> store %jl_value_t* %26, %jl_value_t** %11, align 8
>>> %29 = call %jl_value_t* @jl_apply_generic(%jl_value_t** %4, i32 4)
>>> %30 = load i64, i64* %10, align 8
>>> store i64 %30, i64* %8, align 8
>>> ret %jl_value_t* %29
>>> }
>>> Inside using translator 2
>>>
>>> define double @julia_g_70187(%"#g#143"*, double, double) #0 {
>>> top:
>>> %3 = getelementptr inbounds %"#g#143", %"#g#143"* %0, i64 0, i32 1
>>> %4 = load double, double* %3, align 8
>>> %5 = fmul double %4, %1
>>> ret double %5
>>> }
>>> 1.01
>>> Inside using translator 2 of inside
>>>
>>> define %jl_value_t* @julia_g_70316(%jl_value_t*, double, double) #0 {
>>> top:
>>> %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #2
>>> %ptls_i8 = getelementptr i8, i8* %thread_ptr, i64 -2672
>>> %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
>>> %3 = alloca [6 x %jl_value_t*], align 8
>>> %.sub = getelementptr inbounds [6 x %jl_value_t*], [6 x %jl_value_t*]*
>>> %3, i64 0, i64 0
>>> %4 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 2
>>> %5 = bitcast %jl_value_t** %4 to i8*
>>> call void @llvm.memset.p0i8.i32(i8* %5, i8 0, i32 32, i32 8, i1 false)
>>> %6 = bitcast [6 x %jl_value_t*]* %3 to i64*
>>> store i64 8, i64* %6, align 8
>>> %7 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 1
>>> %8 = bitcast i8* %ptls_i8 to i64*
>>> %9 = load i64, i64* %8, align 8
>>> %10 = bitcast %jl_value_t** %7 to i64*
>>> store i64 %9, i64* %10, align 8
>>> store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
>>> %11 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 5
>>> %12 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 4
>>> %13 = getelementptr [6 x %jl_value_t*], [6 x %jl_value_t*]* %3, i64 0,
>>> i64 3
>>> %14 = bitcast %jl_value_t* %0 to i64*
>>> %15 = load i64, i64* %14, align 8
>>> %16 = getelementptr %jl_value_t, %jl_value_t* %0, i64 1
>>> %17 = bitcast %jl_value_t* %16 to i64*
>>> %18 = load i64, i64* %17, align 8
>>> %19 = bitcast %jl_value_t** %4 to i64*
>>> store i64 %15, i64* %19, align 8
>>> %20 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %21 = getelementptr inbounds %jl_value_t, %jl_value_t* %20, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %21, align 8
>>> %22 = bitcast %jl_value_t* %20 to double*
>>> store double %1, double* %22, align 8
>>> store %jl_value_t* %20, %jl_value_t** %13, align 8
>>> %23 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %24 = getelementptr inbounds %jl_value_t, %jl_value_t* %23, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %24, align 8
>>> %25 = bitcast %jl_value_t* %23 to double*
>>> store double %2, double* %25, align 8
>>> store %jl_value_t* %23, %jl_value_t** %12, align 8
>>> %26 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32
>>> 16)
>>> %27 = getelementptr inbounds %jl_value_t, %jl_value_t* %26, i64 -1,
>>> i32 0
>>> store %jl_value_t* inttoptr (i64 139896322417392 to %jl_value_t*),
>>> %jl_value_t** %27, align 8
>>> %28 = bitcast %jl_value_t* %26 to i64*
>>> store i64 %18, i64* %28, align 8
>>> store %jl_value_t* %26, %jl_value_t** %11, align 8
>>> %29 = call %jl_value_t* @jl_apply_generic(%jl_value_t** %4, i32 4)
>>> %30 = load i64, i64* %10, align 8
>>> store i64 %30, i64* %8, align 8
>>> ret %jl_value_t* %29
>>> }
>>>
>>>
>>>
>>>
>>>
>>>
>>> So in the end, I couldn't find a way within a function to enclose the
>>> parameter α and compile a function which actually treats α as a constant
>>> and optimizes it all the way. However, the ParameterHolder and translator
>>> results are getting pretty close, but I can't seem to get rid of the bounds
>>> checking.
>>>
>>> Does anyone else have a better solution? Or is this supposed to "act
>>> nicer" by default?
>>>
>>
>>
>>
>> --
>> Erik Schnetter <[email protected]>
>> http://www.perimeterinstitute.ca/personal/eschnetter/
>>
>
