Consider inlining fldmod to allow optimizing if an arg is a Const

[NHDaly]

Opening an issue based on this discussion: JuliaMath/FixedPointDecimals.jl#43 (review)

In our code, fldmod isn't inlining. Where we are calling it, one of the arguments is a static compiler Const, and so the division could be optimized away (into a cheaper multiplication and bitshift -- explanation here: JuliaMath/FixedPointDecimals.jl#43 (comment)), but it isn't able to because the non-inlined fldmod presents a function barrier that prevents the compiler from knowing one of the arguments is a Const.

fldmod is such a simple function, so i'm surprised it wasn't inlining anyway. Would it make sense to mark this as @inline in Base? If you think so, i can send a PR to do it.

[vchuravy]

Is this on 1.0 or on 1.X-dev? And do you have a MWE?

fldmod1 is inlined for me on both 1.0 and 1.X-dev (with normal constant values), but only 1.X-dev is doing the constant propagation.

f() = fldmod1(1, 2)

# 1.0.2
julia> @code_typed f()
CodeInfo(
1 1 ─      (Base.checked_sdiv_int)(1, 2)::Int64                 │╻╷╷  fldmod1
  │        (Base.checked_sdiv_int)(1, 2)::Int64                 ││╻╷╷  mod1
  │   %3 = π (1, Int64)                                         ││╻    mod1
  │        (Base.ifelse)(false, 2, %3)::Int64                   │││  
  └──      return (1, 1)                                        │    
) => Tuple{Int64,Int64}

# 1.X-dev
@code_typed f()
CodeInfo(
1 ─     return (1, 1)
) => Tuple{Int64,Int64}

-- edit:
Or did you mean if only one of the arguments is const?

[NHDaly]

Yeah only one of the arguments is const.

[NHDaly]

We talked in person, maybe the better, broader solution is to change the inlining heuristics to be aware of constant propagation.

So, if inlining might not seem worth it on its face, but it allows you to propagate a Const you couldn't do otherwise, that should increase the likelihood of inlining by some degree.

---

Based on some preliminary exploration, maybe that change could be something like adding a check into this function:

julia/base/compiler/optimize.jl

Lines 360 to 361 in 73c8eb4

	function inline_worthy(body::Array{Any,1}, src::CodeInfo, spvals::SimpleVector, slottypes::Vector{Any},
	params::Params, cost_threshold::Integer=params.inline_cost_threshold)

that if any of the arguments to this are Core.Compiler.Consts, maybe you should increase the cost_threshold by some amount?

[JeffreySarnoff]

divrem should be included in this (if it too is not being inlined with const arg[s].