tom.deneau at amd.com
Tue Jun 25 08:00:59 PDT 2013
In HSAIL (when we finally define our own callingConvention and stop using the x64 convention) arguments cannot be passed on the stack anyway.
From: Doug Simon [mailto:doug.simon at oracle.com]
Sent: Tuesday, June 25, 2013 1:43 AM
To: Deneau, Tom
Cc: graal-dev at openjdk.java.net
Subject: Re: stackslot question
On Jun 24, 2013, at 9:07 PM, "Deneau, Tom" <tom.deneau at amd.com> wrote:
> In HSAIL there is not stack register as such but we can declare an array that is private to each workitem
> and reference at positive offsets into it.
> I see there is a bug in our current HSAIL implementation of stack slots which needs to be fixed
> and we are exploring different ways to fix it.
> My question is, assuming a load from a certain stackslot always reads back what was stored to that stackslot,
> (when the load and store both using the same data width)
> is there any requirement on the relative positions of different stackslots in memory,
> for example does it matter whether stackslot -8 is at a higher or lower memory address compared to stackslot -16?
The only time I think stack slot ordering matters is when arguments are passed via the stack. The ordering in that case is defined by the underlying CallingConvention in use. Otherwise, the relative position of stack slots in memory should not matter.
More information about the graal-dev