This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

initial vectorization functionality now in mainline

> The patch is ok.

thanks. tested on powerpc-apple-darwin7.0.0 and i686-pc-linux-gnu, and

New tests results:
on ppc-darwin:
# of expected passes            148
# of expected failures          38
on i686:
# of expected passes            132
# of expected failures          54

1) Most of the xfails on i*86 and x86_64 are due to the new restriction I
added that considers STACK_BOUNDARY. Can something better be done for these
platforms? (should I check against PREFERRED_STACK_BOUNDARY instead?)

2) The current vectorization functionality in mainline is _very_ limited.
It will be significantly improved with the following patches for mainline
that are coming up:
- support for loops with unknown loop bound (or loop bound that doesn't
divide by the vectorization factor)
- support for misaligned accesses
- peeling to force alignment of accesses
- support for vectorization of conditional operations
- support for additional forms of data-references (including
multidimensional arrays)

Description of the vectorizer and it's current status can be found here:
(last updated 4 weeks ago - I'll submit an update shortly)


                      Richard Henderson                                                                                   
                      <>         To:       Dorit Naishlos/Haifa/IBM@IBMIL                                   
                      17/08/2004 00:31         Subject:  Re: [patch] lno branch merge -- vectorizer patch #1              

The patch is ok.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]