[llvm-commits] [PATCH] Adding the Linpack test to the test suite

[Duncan Sands]

On 27/11/12 18:27, Renato Golin wrote:
> On 27 November 2012 17:18, Duncan Sands <baldrick at free.fr> wrote:
>> Unfortunately moving the cx and px initialization down to the 101 length
>> loop a bit below doesn't fix the test :(
>
> Doesn't fix the test == print a different number than the "good one",
> or print two different numbers based on optimization?

Prints two different numbers.

>> Another issue is why globalopt broke a1 into individual fields when it can
>> see that users are accessing off the end of arrays (which is valid as long as
>> they are still indexing into the containing structure).
>
> Not sure what you're talking about... ;)

If you run "opt -internalize -internalize-public-api-list=main" and
"opt -internalize -internalize-public-api-list=main -globalopt" and
compare the global variables, you will see what I mean!

>
> But I'm glad that we're already finding odd behaviour of the optimizer
> with this test!

Indeed!

Ciao, Duncan.

>
> --
> cheers,
> --renato
>
> http://systemcall.org/
>