[LLVMdev] is there a compiler barrier that is effective at codegen level?

[Owen Anderson]

On Dec 22, 2009, at 12:57 PM, Torvald Riegel wrote:

> Here is an example:
> 
>  %160 = add i64 %158, %159                       ; <i64> [#uses=1]
>  %161 = sub i64 %160, %startHash.0               ; <i64> [#uses=5]
>  tail call void asm sideeffect "", "~{dirflag},~{fpsr},~{flags},~{memory}"() 
> nounwind
>  %fp73 = call i8* @llvm.frameaddress(i32 0)      ; <i8*> [#uses=1]
>  %sp74 = call i8* @llvm.stacksave()              ; <i8*> [#uses=1]
>  call fastcc void @Function1(i8* %sp74, i8* %fp73) nounwind
>  %162 = call i32 (i32, ...)* @Function2(i32 0) ; <i32> [#uses=1]
>  tail call void asm sideeffect "", "~{dirflag},~{fpsr},~{flags},~{memory}"() 
> nounwind
> 
> In particular, I would like to prevent modifications to the stack to be put 
> between the calls to Function1 and Function2. Is this possible at all?
> The standard asm compiler barriers don't seem to work, can I use another hack?
> (I could put the calls to Function1 and Function2 into another noinline dummy 
> function and call this one instead, but is there some kind of barrier?)
> 

Putting one of those asm hacks, or a memory fence intrinsic, between the two calls should do the trick.  If it doesn't please file a bug with an example including the input code, the output code, and (preferably) what optimization is messing it up.

--Owen
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 2620 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20091222/48335875/attachment.bin>