Request for reviews (XS): 6843752: missing code for an anti-dependent Phi in GCM

[Tom Rodriguez]

Based on the old code that case definitely seems needed and I believe  
I follow the logic enough to think that it would end up handling it  
the same as the old code.  Looks good.

tom

On May 27, 2009, at 9:31 AM, Vladimir Kozlov wrote:

> Resending.
>
> http://cr.openjdk.java.net/~kvn/6843752/webrev.00
>
> Fixed 6843752: missing code for an anti-dependent Phi in GCM
>
> Problem:
> Changes for 6470497 (part G) incorrectly removed the code
> for anti-dependent PHI pinned below load's 'early' block.
> As result the load placed below this phi into the block with
> lowest frequency and a wrong value is loaded.
> In the test case the load placed inside loop which modifies
> the loaded value to NULL. As result we got NULL exception.
>
> Solution:
> Don't place a load below anti-dependent PHI.
> Added the regression test modified to show the problem
> in latest VM (the loop body executed less time then
> the code before the loop).
>
> Reviewed by:
>
> Fix verified (y/n): y,  bug's test
>
> Other testing:
> JPRT
>