[ccache] Re: trouble with icc

[Robert Anderson]

On 3/13/06, Jason Lunz <lunz at falooley.org> wrote:
>
> rwa000 at gmail.com said:
> > Can anyone help out here?  I am working in an absolutely huge software
> > project and ccache would be a boon if I could just make it work with
> icc.
>
> I gave up on using ccache for icc. afaict, ccache assumes (reasonably)
> that a compiler's output should be the same regardless of whether
> preprocessing and compilation are done as one step or two.



I have independently discovered this, and am pretty horrified that this is
not invariant behavior in icc.

Unfortunately, the enormous icc frontend seems to be filled with magic
> undocumented hacks that turn on various internal workarounds depending
> on the input. The best I could tell, the set of hacks that gets enabled
> is different when you feed it already-preprocessed code.
>
> To see this, you can replace the icc backend (mcpcom) with a wrapper
> that saves a copy of the mcpcom's response file input before invoking
> the real mcpcom. The icc frontend generates a tempfile of this input
> each time it runs mcpcom, and by saving this and looking at it you can
> see what it's really doing. In fact, you may be able to discern from
> this what magic input options are required to get the frontend to always
> compile your code the same way, and that could allow you to use ccache.
>
> As for me, I developed a healthy Fear of icc after investigating this
> far and stopped trying to use it with ccache.


Yeah, I think I give up too.

After spending two weeks recently diagnosing a bug in the floating point
evaluation in icc, I think I am throwing up my hands and migrating to g++.

Thanks for your comments.

Bob