Hi Kelvin,
On Tue, Apr 09, 2019 at 08:15:31AM -0500, Kelvin Nilsen wrote:
> This new patch addresses the code generation problems that were uncovered by
> these failing tests. Additionally, this new patch corrects some of the
> expected instruction counts for certain previously existing regression tests
> on certain targets to adjust for changes in the generated code.
Is the newly generated code better though? Worse? More expected? It isn't
clear to me.
> This new patch has been bootstrapped and tested without regressions on
> powerpcle-unknown-linux (both P8 and P9) and on powerpc-linux (P7 and P8,
> both -m32 and -m64).
powerpcle-linux is a very different configuration than the powerpc64le-linux
you mean (powerpc64-linux and powerpc-linux are biarch to each other, but
the LE variants are not). Very minor... It's just that powerpcle-linux
makes me cringe.
> * gcc.target/powerpc/fold-vec-extract-short.p8.c: Likewise.:
(stray colon)
> +// P8 LE variable offset: rldicl, subfic, sldi, mtvsrd, xxpermdi, vslo,
> mfvsrd, sradi, rlwin, (extsb)
rlwinm
> -/* { dg-final { scan-assembler-times {\mrlwinm\M} 2 { target lp64} } } */
> +/* { dg-final { scan-assembler-times {\mrlwinm\M} 4 { target lp64} } } */
Put a space after lp64 while you're here? Or remove the one before target,
whichever you like best.
I'm a bit worried that these tests change instruction counts so often, what
that means for maintainability. But, okay for trunk, and for backports (but
please make sure they generate the correct counts for all targets there :-/ )
Segher
