eliminate constant dead branches earlier?

[StefanKarpinski]

Currently, we have this:

julia> const debug = false
false

julia> function foo()
           if debug
               println("Debugging...")
           end
           1 + 2
       end
foo (generic function with 1 method)

julia> @code_llvm foo()

define i64 @julia_foo17924() {
top:
  %0 = alloca [4 x %jl_value_t*], align 8
  %.sub = getelementptr inbounds [4 x %jl_value_t*]* %0, i64 0, i64 0
  store %jl_value_t* inttoptr (i64 4 to %jl_value_t*), %jl_value_t** %.sub, align 8
  %1 = load %jl_value_t*** @jl_pgcstack, align 8, !dbg !876
  %2 = getelementptr [4 x %jl_value_t*]* %0, i64 0, i64 1, !dbg !876
  %.c = bitcast %jl_value_t** %1 to %jl_value_t*, !dbg !876
  store %jl_value_t* %.c, %jl_value_t** %2, align 8, !dbg !876
  store %jl_value_t** %.sub, %jl_value_t*** @jl_pgcstack, align 8, !dbg !876
  %3 = getelementptr [4 x %jl_value_t*]* %0, i64 0, i64 2
  store %jl_value_t* null, %jl_value_t** %3, align 8
  %4 = getelementptr [4 x %jl_value_t*]* %0, i64 0, i64 3
  store %jl_value_t* null, %jl_value_t** %4, align 8
  %5 = load %jl_value_t** %2, align 8, !dbg !877
  %6 = getelementptr inbounds %jl_value_t* %5, i64 0, i32 0, !dbg !877
  store %jl_value_t** %6, %jl_value_t*** @jl_pgcstack, align 8, !dbg !877
  ret i64 3, !dbg !877
}

It would be nice if the fact that the if debug branch can't possibly happen were detected early enough to avoid allocation of a GC frame, allowing the branch to get fully eliminated.

[ArchRobison]

Could constant values be propagated by extending the type-inference framework? Constant propagation and type inference in Julia are similar forward-flow problems.

[JeffBezanson]

dup of #5560