Re: compile/920501-18.c and -fssa

[Brad Lucier <lucier at math dot purdue dot edu>]

Mark Mitchell wrote:

> >>>>> "Richard" == Richard Kenner <kenner@vlsi1.ultra.nyu.edu> writes:
> 
>     Richard> I think there's something wrong with the transformation
>     Richard> of the loop.  Could somebody who understands the SSA
>     Richard> stuff please verify?
> 
> There are known issues with SSA, basically because the design of
> RTL-generation involves generating overly low-level constructs early
> in the game.  Therefore, I wouldn't suggest we spend effort on this at
> the moment.

On the Alpha 21264, -fssa reorders floating-point code in a manner that
happens to work around a difficulty in gcc's register allocator for IEEE
floating-point arithmetic; see a series of e-mails in

http://gcc.gnu.org/ml/gcc/2000-03/

with the (rather unfortunate) title

-fssa kicks butt on alphaev6 ieee floating point code

On alphaev6-unknown-linux-gnu, then gcc.c-torture/compile/920501-18.c is
the *only* test that fails with -fssa and does not fail with other
compile options.  So please, let's not abandon it yet.

> In fact, I think it's probably a waste of time to run
> with -fssa in the testsuite, for now.

I disagree.

Brad Lucier