[llvm-commits] [llvm] r61969 - in /llvm/trunk: include/llvm/Transforms/IPO/InlinerPass.h include/llvm/Transforms/Utils/InlineCost.h lib/Transforms/IPO/InlineAlways.cpp lib/Transforms/IPO/InlineSimple.cpp lib/Transforms/IPO/Inliner.cpp lib/Transforms/Utils/InlineCost.cpp

[Dale Johannesen]

On Jan 9, 2009, at 1:03 AMPST, Duncan Sands wrote:

> Hi Dale,
>
>> +    // Don't inline a callee with dynamic alloca into a caller  
>> without them.
>> +    // Functions containing dynamic alloca's are inefficient in  
>> various ways;
>
> would they be less inefficient if codegen output all fixed size  
> allocas before
> any variable sized alloca's?  Then all fixed size allocas will be at  
> a constant
> offset from the frame pointer.  If there is no frame pointer then I  
> guess it
> should be inverted: dynamic allocas first, then fixed size ones, so  
> the fixed
> size allocas are at a constant offset from the stack pointer.  But  
> perhaps
> there are other inefficiencies beyond having to compute the  
> addresses of
> stack variables?

Calls that push parameters on the stack have to decrement the stack  
pointer
dynamically instead of allocating a fixed-offset parameter area in the  
prolog.
This is very noticeable on x86-32.

> Ciao,
>
> Duncan.
>
> PS: Ada uses plenty of variable sized allocas.  So I'm interested in  
> minimizing
> the impact of them.  Also, rather than tweaking the inliner like  
> this, mightn't
> be better to reduce the impact of using variable sized allocas to  
> the point where
> inlining them is acceptable?