[aarch64-port-dev ] A rather large patch to aarch64.ad

[Andrew Dinn]

On 25/10/13 17:37, Andrew Haley wrote:
> This is a substantial rewrite of the AArch64 patterns for arithmetic
> and memory access.  The idea is to make use of the more complex
> AArch64 instructions wherever possible, so that e.g. we use UBFX
> rather than a series of shifts and masks, and LDR Xd, [Xn, Wn, sxtw #3]
> for array loads.
> 
> The patch itself is hard to follow because of the reorganization I had
> to do.  I apologize for that.  However, the resulting aarch64.ad isn't
> so bad.
> 
> I have checked many of these patterns and HotSpot really does generate
> the fast AArch64 instructions.
> 
> Comments?

Apologies for the lack of comment. The new rules look good but I have
not yet tried them. I have been holding off pulling these changes down
while I try to pin down the register pressure problem thrown up by
eclipse -- just in case the pulled changes make the bug go underground.
I think I have pinned down exactly where things go wrong but am still
not clear exactly why.

However, since it takes almost 1/2 hour to reach the point of error I
think I will try out a build which includes the new rules. If the bug
still remains then at least I will be able to restart a debug session
somewhat faster.

regards,


Andrew Dinn
-----------