llvm dev - Mar 2011 - [LLVMdev] Kinda noob questions

On Mar 14, 2011, at 5:56 PM, John Criswell wrote:
> On 3/14/11 7:52 PM, Judison wrote:
>> 
>> 2) does the LLVM optimizes this:
>> 
>> ; this code (arg -> var) is very similar to what gcc generates
>> define void @foo(i32 %arg0) {
>> entry:
>>     %var0 = alloca i32
>>     store i32 %arg0, i32* var0
>>     ...
>>     %x = load i32* %var0
>>     ....
>>     ; never more stores to %var0 or pass it to any function
>> }
>> 
>> to something like:
>> 
>> define void @foo(i32 arg0) {
>> entry:
>>     ; no alloca
>>     ....
>>     %x = %arg0
>>    ....
>> }
>> 
>> does it???
> 
> I suspect that mem2reg plus a few additional optimizations will do this for
you.
mem2reg is sufficient.  There's also no limit on the number of loads
and stores you can do without disturbing the optimization;  just
don't do anything too opaque with the address.

John.

Thank you John and John :P

these optimizations (mem2reg and the "few additional" ones) I have to
enable
then or something like this?? (I compile the .ll to .o with the sequence
llvm-as, llc, as)

Yet about dynamic stack allocation, what is (generally) better? to pre alloc
everything at start, or let it be?

Imagine this pseudo-code

while (x) {
  int b = 0;
  ...
}

using alloca where b is declared, it will grow the stack dynamically, but it
will shrink it at } or in llvm's words in the end of a block where b is not
accessible anymore? Can I force its deallocation??

Thanks again

On Mon, Mar 14, 2011 at 10:23 PM, John McCall <rjmccall at apple.com>
wrote:
> On Mar 14, 2011, at 5:56 PM, John Criswell wrote:
> > On 3/14/11 7:52 PM, Judison wrote:
> >>
> >> 2) does the LLVM optimizes this:
> >>
> >> ; this code (arg -> var) is very similar to what gcc generates
> >> define void @foo(i32 %arg0) {
> >> entry:
> >>     %var0 = alloca i32
> >>     store i32 %arg0, i32* var0
> >>     ...
> >>     %x = load i32* %var0
> >>     ....
> >>     ; never more stores to %var0 or pass it to any function
> >> }
> >>
> >> to something like:
> >>
> >> define void @foo(i32 arg0) {
> >> entry:
> >>     ; no alloca
> >>     ....
> >>     %x = %arg0
> >>    ....
> >> }
> >>
> >> does it???
> >
> > I suspect that mem2reg plus a few additional optimizations will do
this
> for you.
>
> mem2reg is sufficient.  There's also no limit on the number of loads
> and stores you can do without disturbing the optimization;  just
> don't do anything too opaque with the address.
>
> John.



-- 
Judison
judison at gmail.com

"O ignorante que procura se instruir é como um sábio; o sábio que fala sem
discernimento se assemelha a um ignorante." Imam Ali (as)
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20110314/8497cd56/attachment.html>

On Mar 14, 2011, at 6:44 PM, Judison wrote:
> Thank you John and John :P
> 
> these optimizations (mem2reg and the "few additional" ones) I
have to enable then or something like this?? (I compile the .ll to .o with the
sequence llvm-as, llc, as)
Look into 'opt'.
> Yet about dynamic stack allocation, what is (generally) better? to pre
alloc everything at start, or let it be?
> 
> Imagine this pseudo-code
> 
> while (x) {
>   int b = 0;
>   ...
> }
> 
> using alloca where b is declared, it will grow the stack dynamically, but
it will shrink it at } or in llvm's words in the end of a block where b is
not accessible anymore? Can I force its deallocation??
Unless you're actually reliant on dynamic allocation — e.g. you don't
know the size statically, or you're intending to repeatedly allocate memory
— you should always allocate in the entry block.  LLVM is reasonably good at
re-using stack slots when it can.

John.