C[omp]ute

From: andrew cooke <andrew@...>

Date: Fri, 15 May 2015 11:38:45 -0300

Note that this is LLVM IR assembly, not x64.  But you can access even very
specific instructions, like PCLMULQDQ 
http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20120528/143720.html

First, you need to build Julia from trunk with LLVM 3.6 (instructions from
https://groups.google.com/d/msg/julia-users/PffilD_uaeg/hkQh5U7KE2QJ which
really has everything you need to know - thanks Isiah)

  git clone git://github.com/JuliaLang/julia.git julia-llvm-3.6
  cd julia-llvm-3.6
  echo "LLVM_VER=3.6.0" > Make.user
  make -j 3

You can then check that:

  pushd usr/bin
  ./llc -version
  popd

Now, test assembly inside julia:

  f(x::Int64,y::Int64) = Base.llvmcall("""
    %3 = add i64 %0, %1
    ret i64 %3""",
    Int64,
    Tuple{Int64, Int64},
    x, y)
  f(1, 2)

which should give 3.

Similarly

  f() = Base.llvmcall("""
    %1 = add i64 1, 2
    ret i64 %1""",
    Int64,
    Tuple{})

(I have no idea why that needs to be %1 rather than %3, but %3 gives an
error).

Trying to get beyond this, I tried to print a string.  But the reference
manual (link below) uses "declare", which llvmcall doesn't seem to like.  And
I can't get literal strings to work (see "array constants" in the manual,
which says c"hello world" should be ok).

For example,

  g() = Base.llvmcall("""
    call i32 (i8*, ...)* @printf(i8* c"hello world\0")
    ret""",
    Void, Tuple{})

gives "error: expected string" right where the string is, while trying to
define it previously, with something like "@s = ..." gives "error: expected
insruction opcode".


So I tried to look at the output from code_llvm, but that's not terribly
helpful either:

julia> function h()
         print("hello world")
       end
h (generic function with 1 method)

julia> code_llvm(h, ())

define %jl_value_t* @julia_h_44607() {
top:
  %0 = alloca [4 x %jl_value_t*], align 8
  %.sub = getelementptr inbounds [4 x %jl_value_t*]* %0, i64 0, i64 0
  %1 = getelementptr [4 x %jl_value_t*]* %0, i64 0, i64 2
  %2 = bitcast [4 x %jl_value_t*]* %0 to i64*
  store i64 4, i64* %2, align 8
  %3 = getelementptr [4 x %jl_value_t*]* %0, i64 0, i64 1
  %4 = bitcast %jl_value_t** %3 to %jl_value_t***
  %5 = load %jl_value_t*** @jl_pgcstack, align 8
  store %jl_value_t** %5, %jl_value_t*** %4, align 8
  store %jl_value_t** %.sub, %jl_value_t*** @jl_pgcstack, align 8
  store %jl_value_t* null, %jl_value_t** %1, align 8
  %6 = getelementptr [4 x %jl_value_t*]* %0, i64 0, i64 3
  store %jl_value_t* null, %jl_value_t** %6, align 8
  %7 = load %jl_value_t** inttoptr (i64 139727723511144 to %jl_value_t**),
  align 8
  %8 = icmp eq %jl_value_t* %7, null
  br i1 %8, label %err, label %ok

err:                                              ; preds = %top
  call void @jl_undefined_var_error(%jl_value_t* inttoptr (i64 139727711444264
  to %jl_value_t*))
  unreachable

ok:                                               ; preds = %top
  store %jl_value_t* %7, %jl_value_t** %1, align 8
  store %jl_value_t* inttoptr (i64 139727767335888 to %jl_value_t*),
  %jl_value_t** %6, align 8
  %9 = call %jl_value_t* @jl_apply_generic(%jl_value_t* inttoptr (i64
  139727730966960 to %jl_value_t*), %jl_value_t** %1, i32 2)
  %10 = load %jl_value_t*** %4, align 8
  store %jl_value_t** %10, %jl_value_t*** @jl_pgcstack, align 8
  ret %jl_value_t* %9
}

Which doesn't help!


Finally, documentation:

  * There's an exmaple involving tuples in test/llvmcall.jl

  * The structure of the call is given in src/ccall.cpp as
    llvmcall(ir, (rettypes...), (argtypes...), args...)

  * The language ref for 3.6 is at
    http://llvm.org/releases/3.6.0/docs/LangRef.html

  * I can't find anything on x86 intrinsics, however.



Andrew