Re: [split] Add x86 support

[Richard Henderson <rth at redhat dot com>]

On 09/28/2009 04:51 PM, Ian Lance Taylor wrote:
> The plan is not to use thunk-like objects.  I agree that in the
> general case that can not work, because you can't know how many
> parameter bytes to copy on to the new stack.  The plan is for the
> linker to actually modify the size of the stack space request in
> split-stack code which calls non-split-stack code.

Interesting.

> I considered this sort of approach, and my main concern is that this
> increases the performance penalty.  Now every function uses a frame
> pointer, and, less significantly, the overhead is no longer
> lea/cmp/jle, it is movq/call/lea/cmp/jle/ret.

Perhaps you can mix both options, by having different entry points into 
the support library.

The stack realignment solution does have the advantage that you don't 
have to invoke copy constructors for those arguments that need them. 
You could then choose between the two alternatives based on looking at 
the actual arguments to the callee.


r~