[cfe-commits] r81602 - in /cfe/trunk: lib/CodeGen/CGBlocks.cpp lib/CodeGen/CGDecl.cpp lib/CodeGen/CGExpr.cpp test/CodeGen/blocks-aligned-byref-variable.c

[Chris Lattner]

On Sep 13, 2009, at 1:14 PM, Mike Stump wrote:

> On Sep 12, 2009, at 12:54 PM, Chris Lattner wrote:
>> On Sep 12, 2009, at 11:58 AM, Mike Stump wrote:
>>
>>> On Sep 11, 2009, at 7:44 PM, Anders Carlsson wrote:
>>>> URL: http://llvm.org/viewvc/llvm-project?rev=81602&view=rev
>>>> Log:
>>>> Add support for __block variables with alignment greater than
>>>> __alignof(void *).
>>>
>>> Unfortunately, we also need a sema error for alignment larger than  
>>> the
>>> minimum of the maximal stack alignmint and the alignment of malloc  
>>> on
>>> the system.  For darwin, I think that means 16.  Also, ordinary  
>>> stack
>>> decls that request larger than the maximal stack alignment the  
>>> system
>>> conspires to allow need a sema error as well.  In time, people might
>>> include the required alignment in the block and a flag to request  
>>> more
>>> than the standard align in the blocks runtime, but until then...   
>>> I've
>>> added a FIXME.
>>
>> Why?
>
> Why what?  Why do we need an error for going past maximal supported  
> alignment, because it won't be honored.  Why warn for going past the  
> alignment of malloc, because the runtime will dealign a __block  
> variable to just malloc align[1] during a copy of the __block  
> variable.  In either case, this isn't what the user wanted.  They  
> actually wanted a variable that is aligned to the stated boundary.
>
>> Codegen should just set the right alignment on the alloca.  The  
>> LLVM backend will realign the stack as needed.
>
> You say that as if it were true:

It is true, we just don't support alignments of 512M *anywhere*.  If  
you want to solve the greater problem, that's fine, but this isn't  
__block specific at all.

-Chris

>
> void bar(void *);
>
> void foo() {
>  int i __attribute__((aligned(1024*1024*512)));
>  i = 42;
>  bar(&i);
> }
>
> clang -S t.c -o -
>
> _foo:
> 	pushq	%rbp
> 	movq	%rsp, %rbp
> 	subq	$16, %rsp
>
> ?  16 won't cut it here.  Anyway, __block variables can be allocate  
> by other than the llvm stack allocator.  They can be allocated when  
> the runtime copies them.
>
>
> 1 - To be verified when one writes the check.