Int vs float comparisons in Julia

[simonbyrne]

The first commit (a5de252) implements Int64/UInt64 vs Float32/Float64 comparisons in pure Julia, removing the C++ code from src/intrinisics.cpp. From what I can tell, this should emit identical LLVM code as before for Float64; for Float32, the methods should now work without promotion to Float64.

The second commit (62828b7) makes a small modification to fix #9021. Any comparisons with constant integers (e.g. x > 0) should now be inlined into a single float comparison. Is there anyway I could craft this into a test?

[JeffBezanson]

Oo, nice code reduction.

[JeffBezanson]

Maybe we should declare these @inline? @vtjnash what do you think?

[StefanKarpinski]

This is very nice.

[StefanKarpinski]

Declaring @inline seems like a good idea since that's effectively what using an intrinsic does.

[JeffBezanson]

To be more specific, while these functions are inlineable as-is, I worry that a large function with many calls to these could hit a size threshold that prevents further inlining.

[vtjnash]

I generally recommend against forcing inlining. Our current heuristic is pretty generous already, so when you hit that limit you probably really want to stop inlining. One case where that might be false is when you have a large if statement that you know will rarely be triggered and the speed of that code is highly critical. Of the user has redefined some of these internal functions to be large, it is possible that we don't actually want to be inlining here at all. In addition to the size penalty in the AST and resulting code, my understanding is that some of the llvm passes are O(n^2) in the lines of code, so it is best not to make things excessively large

[simonbyrne]

What's the call here: to @inline or not to @inline?

[eschnett]

I tend to use @inline only if I know that inlining will expose further optimization opportunities, e.g. evaluating an if condition at compile time, of if the function is very small (just a wrapper around another function call). Otherwise, it's usually best to leave it to the compiler.

How many machine instructions are generated from the comparison functions? If it's fewer than for a function call, then we should use @inline.

[JeffBezanson]

Can always add it later :)