[PATCH] Allocate stack storage for .block_descriptor and captured self.

[Adrian Prantl]

On Mar 6, 2013, at 1:47 PM, John McCall <rjmccall at apple.com> wrote:

> On Mar 1, 2013, at 10:15 AM, Adrian Prantl <aprantl at apple.com> wrote:
>> On Feb 28, 2013, at 11:20 AM, John McCall <rjmccall at apple.com> wrote:
>>> On Feb 28, 2013, at 11:12 AM, Eric Christopher <echristo at gmail.com> wrote:
>>>> On Thu, Feb 28, 2013 at 11:06 AM, John McCall <rjmccall at apple.com> wrote:
>>>> On Feb 27, 2013, at 3:17 PM, Eric Christopher <echristo at gmail.com> wrote:
>>>>> On Wed, Feb 27, 2013 at 11:49 AM, John McCall <rjmccall at apple.com> wrote:
>>>>> On Feb 27, 2013, at 11:42 AM, Eric Christopher <echristo at gmail.com> wrote:
>>>>>> On Wed, Feb 27, 2013 at 11:39 AM, Adrian Prantl <aprantl at apple.com> wrote:
>>>>>> 
>>>>>> On Feb 27, 2013, at 11:31 AM, John McCall <rjmccall at apple.com> wrote:
>>>>>>> Okay, you're saying that the value is actually no longer live at all at this point in the function?  It seems reasonable to lose track of the debug info then, although we should be leaving behind a marker in the DWARF that says the value is unavailable.
>>>>>>> 
>>>>>>> If we want to make stronger guarantees in -O0 for purposes of debugging — and I think that's reasonable — then throwing the value in an alloca is acceptable.
>>>>>> 
>>>>>> To clarify: Are you suggesting to only generate the alloca at -O0, or are you comfortable with it as it is?
>>>>>> 
>>>>>> If the value isn't live past that spot I'm more comfortable with dropping the debug info there rather than changing the generated code to keep the value live through the end of the function.
>>>>> 
>>>>> Purely out of attachment to the principle that debug info shouldn't change the code?
>>>>> 
>>>>> 
>>>>> Pretty much.
>>>>> 
>>>>> Not losing information has intrinsic value in a debug build.  If we need to emit slightly different code in order to force a value to stay live and thus improve the debugging experience, then so be it.
>>>>> 
>>>>> 
>>>>> Agreed that making the experience better is desirable, but it can make debugging a problem more difficult if the code changes when you turn on debugging symbols.
>>>> 
>>>> Ah, I see your point;  not doing the alloca could slide stack frames around.
>>>> 
>>>> Alright, I agree with emitting it in all -O0 builds.
>>>> 
>>>> Thought if optimization should fix it then perhaps all builds? :)
>>> 
>>> I don't see any point in creating it just for mem2reg to trivially destroy. :)
>>> 
>>>> That said I'll remove the objection to the allocas. We'll need to fix the alloca problem at some point, but making poor Adrian do it right now for this bug when we've got other workarounds already in the source base seems a bit mean.
>>> 
>>> Well, if the value really isn't live anymore, then I'm not sure what the supposed alloca problem is, other than needing to leave breadcrumbs to say that the value isn't available at this point in the function.  We definitely don't want regalloc to be keeping values live just for debug info!
>> 
>> FYI: this is what the patch looks like if output the alloca only at -O0.
> 
> +  if (CGM.getCodeGenOpts().OptimizationLevel == 0)
> +    blockAddr = Builder.CreateLoad(blockAddr);
> 
> No.  The kind of value in LocalDeclMap should not vary according to
> target optimization level.

Following yours and Eric’s suggestions I got rid of most OptmiziationLevel-dependent code.

> What you should do is:
> - under -O0, always emit an alloca for the block context parameter and store
>   the block context into it, and 
> - under -g, tie the debug info the block context parameter to the alloca, if it
>   exists, and otherwise to the raw argument value.

That works well for the block context parameter.
> 
> +      if (CGM.getCodeGenOpts().OptimizationLevel == 0) {
> +	llvm::Value *selfAddr =
> 
> Don't do this.  If you need to re-indent, re-indent.

Tabs vs spaces. I finally fixed my .emacs.

> +// RUN: %clang_cc1 %s -O0 -fblocks -triple x86_64-apple-darwin -emit-llvm -o - | FileCheck %s
> 
> I'm confused, because that's the default.

You’re correct, the -O0 is redundant.

> +// RUN: %clang -fblocks -S -g -fverbose-asm -triple x86_64-apple-darwin -o - %s | FileCheck %s
> 
> Like Eric mentioned, tests that need to check backend output shouldn't go
> in the clang test suite.  Change the test to check IR output or move it to
> LLVM.


I split the benchmark in two: One part tests ObjC -> llvm IR generation and another one tests llvm IR -> Dwarf.

Thanks for the many comments so far,
-- adrian

-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Test-that-we-emit-a-DW_AT_location-for-self-captured.patch
Type: application/octet-stream
Size: 17636 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20130307/c18c1c4f/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Allocate-stack-storage-for-.block_descriptor-and-cap.patch
Type: application/octet-stream
Size: 10319 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20130307/c18c1c4f/attachment-0001.obj>