[aarch64-port-dev ] RFR: JDK-8203041 : AArch64: fix overflow in immediate cmp/cmn instruction

[Joshua Zhu]

Hi,

Thanks Andrew for detailed explanation. It is helpful for me.
I submit http://cr.openjdk.java.net/~zyao/8203041/webrev.02/ for review.
I think there will be another change to cover all rscratch alias issue mentioned by Andrew. Therefore I do not touch it.

Best Regards,
Joshua

-----Original Message-----
From: Andrew Haley <aph at redhat.com> 
Sent: Monday, May 21, 2018 7:45 PM
To: Joshua Zhu <Joshua.Zhu at arm.com>; Dmitrij Pochepko <dmitrij.pochepko at bell-sw.com>; aarch64-port-dev at openjdk.java.net
Cc: nd <nd at arm.com>
Subject: Re: [aarch64-port-dev ] RFR: JDK-8203041 : AArch64: fix overflow in immediate cmp/cmn instruction

Hi,

On 05/21/2018 08:52 AM, Joshua Zhu wrote:
> Thanks for your review.
> 
> In my patch I decided not to replace cmp with subs at callsite (in
> generate_large_array_equals()) because imm cmp has accidentally been 
> misused for serval times such as
> https://bugs.openjdk.java.net/browse/JDK-8184900
> https://bugs.openjdk.java.net/browse/JDK-8161190

That's a fair point.  I have made the same mistake myself.

> I added an assert to ensure that for large imm all callsites do not 
> use rscratch1.

Does that help at all?  rscratch1 might be live, even if it is not use in the cmp instruction.

> But your suggestion makes sense. Rscratch1 may imports more confusion. 
> Let's replace cmp with subs at callsite when this assertion failure 
> happens.  The only concern is that this depends on engineers be aware 
> of immediate bits limitation for imm cmp instruction.

Well, yes.  As far as I can see it's a limitation of the instruction set.  I can't see any reasonable way to fix it.

Andrew.