llvm dev - Mar 2010 - [LLVMdev] Inlining into a big function

Hi,

The InlineCostAnalyzer has contained this since it was created:

+  // Don't inline into something too big, which would make it bigger. 
Here, we
+  // count each basic block as a single unit.
+  //
+  InlineCost += Caller->size()/20;

It discourages inlining into very big functions. (100 basic blocks in the caller
= 1 instruction in the callee.) Does anybody remember the rationale behind that
heuristic?

I am currently looking at a very big function with enough basic blocks that the
heuristic effectively disables inlining at -Os. Even for trivial functions that
would shrink the size of the function!

/jakob

On Mar 8, 2010, at 7:17 PM, Jakob Stoklund Olesen wrote:
> Hi,
> 
> The InlineCostAnalyzer has contained this since it was created:
> 
> +  // Don't inline into something too big, which would make it bigger. 
Here, we
> +  // count each basic block as a single unit.
> +  //
> +  InlineCost += Caller->size()/20;
> 
> It discourages inlining into very big functions. (100 basic blocks in the
caller = 1 instruction in the callee.) Does anybody remember the rationale
behind that heuristic?
> 
> I am currently looking at a very big function with enough basic blocks that
the heuristic effectively disables inlining at -Os. Even for trivial functions
that would shrink the size of the function!
No, it seems reasonable to zap it.  The only reason I can imagine for something
like this is to prevent compile time explosion.

-Chris