Extend llvmcall with support for entering declarations.

[maleadt]

This is a copy of PR #8673 -- I accidentally created the old PR from my master branch, which isn't very clean.

This PR adds support to llvmcall for inserting required declarations in the module. This is useful when you want to llvmcall intrinsics, because LLVM refuses to parse IR with undeclared identifiers (for example, see this thread).

The idea behind this solution is to allow passing a tuple (decls, ir)as first argument to llvmcall, where currently only an IR string or a pointer to a LLVM function is allowed. The decls IR is then pasted before the function llvmcall generates. In order to avoid duplicate declarations, which LLVM does not like, a global set keeps track of the declarations added to the current module.

Example usage:

function floor(x::Float64)
  llvmcall(
    ("""declare double @llvm.floor.f64(double)""",
     """%2 = call double @llvm.floor.f64(double %0)
        ret double %2"""),
    Float64, (Float64,), x)
end

As mentioned in issue #8308 I also tried other approaches, such as automatically parsing identifiers in the IR, but that turned out pretty fragile.

[simonbyrne]

It would be great to have something like this to be able to use LLVM intrinsics from with llvmcall. @Keno You mentioned on another thread about making declarations automatic?

[SimonDanisch]

Is this related?
It seems, that llvmcall does try to declare the type correctly, but it somehow doesn't get through with it.
Bug, not implemented or my fault?

immutable Matrix4x4
    i_1::Float64
    i_2::Float64
    i_3::Float64
    i_4::Float64
    i_5::Float64
    i_6::Float64
    i_7::Float64
    i_8::Float64
    i_9::Float64
    i_10::Float64
    i_11::Float64
    i_12::Float64
    i_13::Float64
    i_14::Float64
    i_15::Float64
    i_16::Float64
end

function mulllvm(a::Matrix4x4, b::Matrix4x4)
    Base.llvmcall("""
        %ptr = getelementptr <4 double>, %Matrix4x4* %0, i64 0, i64 0
        %3 = load <4 x double>, <4 double>* %ptr
        ret <4 x double> %3
        """, NTuple{4, Float64},
        (Matrix4x4, Matrix4x4),
        a,b)
end
const x = Matrix4x4(rand(16)...)
println(mulllvm(x,x))
#yields:
ERROR: LoadError: error compiling mulllvm: Failed to parse LLVM Assembly: 
julia: <string>:2:56: error: expected ')' at end of argument list
define <4 x double> @"julia_mulllvm_426531"(%Matrix4x4 = type { double, double, double, double, double, double, double, double, double, double, double, double, double, double, double, double } ,%Matrix4x4 = type { double, double, double, double, double, double, double, double, double, double, double, double, double, double, double, double } ) {
                                                       ^

 in include_from_node1 at loading.jl:129
while loading C:\Users\Sim\.julia\v0.4\playground.jl\llvmcall.jl, in expression starting on line 30

[ihnorton]

This would be really helpful for debugging codegen issues.

[tkelman]

Closing in favor of #11604, thanks to @vchuravy for picking up this work