[llvm-commits] [llvm] r85346 - in /llvm/trunk: include/llvm/CodeGen/Passes.h lib/CodeGen/BranchFolding.cpp lib/CodeGen/BranchFolding.h lib/CodeGen/IfConversion.cpp lib/CodeGen/LLVMTargetMachine.cpp lib/Target/ARM/ARMTargetMachine.cpp

[Evan Cheng]

On Oct 28, 2009, at 8:55 PM, Chris Lattner <clattner at apple.com> wrote:

>
> On Oct 28, 2009, at 7:56 PM, Sandeep Patel wrote:
>
>> On Wed, Oct 28, 2009 at 8:24 PM, Chris Lattner <clattner at apple.com>  
>> wrote:
>>> On Oct 28, 2009, at 11:18 AM, Evan Cheng wrote:
>>>>
>>>> This is tricky. I don't think there is a *right* answer. On x86, I
>>>> don't believe the extra branch matters. To me, tail merge for 1
>>>> instruction tails is not saving much code size and it's potentially
>>>> bad for performance. We should disable it except for -Os.
>>>>
>>>> I don't know if we want to differentiate between -O2 and -O3 here.
>>>> We either do tail merging or we disable tail merging. Unless we can
>>>> show it does hurt performance I'd prefer to leave it along.
>>>
>>> I agree with Evan: the only case that should change is tail merge  
>>> of 1
>>> instruction, which should be enabled at -Os only, and only when it
>>> actually reduces # instructions.  O2 and O3 should have the same
>>> policy IMO.
>>
>> On ARM, a simple BX_RET can be folded into any conditional branch  
>> that
>> points at it, so tail merging even 1 instruction may help.

I think that's if-coverter's job.

Evan
>>
>> PPC also has conditional returns, right?
>
> Yes on both counts.  To be perfectly clear, I only meant when there  
> would also have to be unconditional branch after the single  
> instruction.  If it is the last branch (e.g. a no return call,  
> return, etc) then go for it :)
>
> -Chris