Re: [PATCH 4/4] Initial version of RTL frontend

[Jeff Law <law at redhat dot com>]

On 05/10/2016 08:13 AM, David Malcolm wrote:
> On Wed, 2016-05-04 at 16:49 -0400, David Malcolm wrote:
> [...snip...]
> > I wrote this by compiling a test.c with -fdump-rtl-all on
> > x86_64-pc-linux-gnu, and then attempting to load the state at each
> > pass, and then trying to run just one pass, and fixing bugs until the
> > dump output from each pass was the same as when running the pass from
> > cc1.  I've only done this with a small subset of passes, and with a
> > very
> > simple test case, so I'm sure there are plenty of bugs and "x86_64
> > -isms"
> > remaining.
>
> [...snip...]
>
> On the subject of "x86_64-isms", a couple of issues I ran into when
> testing with --target=aarch64-linux-gnu:
>
> * roundtrip.exp fails due to different register names so e.g. this from
> x86_64:
>
>      (reg:SI 5 di [ i ]))
>
>   becomes this after roundtripping on aarch64:
>
>      (reg:SI 5 x5 [ i ]))
>
>   (i.e. "di" vs "x5" for register 5).
>
> * unknown modes e.g. this from x86_64:
>
>     (reg:CCGC 17 flags)
>
>   fails on aarch64 due to the lack of a "CCGC" mode.
Fundamentally all this stuff is target dependent.  Trying to read in an 
x86-64 generated dump into an aarch64 targeted compiler is a waste of 
time/effort.  THe number of things you're going to stumble over are 
innumerable.

I think we lay down the law that RTL dumps are target specific.    There 
probably should be some kind of metadata emitted in the dump file which 
identifies the target and if we try to read in the wrong stuff, we get a 
nice error.

jeff