[llvm-commits] [llvm-gcc-4.2] r63123 - in /llvm-gcc-4.2/branches/Apple/Dib/gcc: cgraphunit.c ipa-inline.c passes.c tree-pass.h

[Chris Lattner]

On Jan 27, 2009, at 12:57 PM, Dale Johannesen wrote:
>>> Is this some general performance issue, or just that some specific
>>> code that's been extensively tuned to work well with gcc's inliner
>>> doesn't work as well now?  The latter would be expected
>>
>> I considered it a general performance issue that just happen to be
>> exposed by a particular test case.
>
> Do you have evidence for this?   I've looked at it quite a bit, and
> concluded that llvm-gcc's inliner does less inlining on average than
> gcc's, but that this is not, in general, a problem:  in other words,
> benchmarks that are better and worse are about equal in number.
> (And performance being roughly equal, smaller code size is to be
> preferred.)

I can give two easy examples.  1: SROA is not promoting aggregates in  
all cases that it could.  This is particularly a problem for the "by  
value struct passing" stuff in the ABI lowering code of llvmgcc.  2:  
llvm's optimizer doesn't do C++ copy constructor elision.

Both of these mean that if the GCC inliner runs before tree -> LLVM,  
that you'll get the optimizations and performance will be good.   
Running tree -> LLVM before the LLVM inliner will prevent these from  
happening.

-Chris