[commit: ghc] simd: Make sure the allocator aligns memory on a 16-byte boundary. (08ba8f7)

[Johan Tibell]

Is this going into 7.4? Does this mean all heap objects now use 4 words
(e.g. Int is now twice as big)?

On Tue, Nov 1, 2011 at 10:22 AM, Geoffrey Mainlain <gmainlan at galois.com>wrote:

> Repository : ssh://darcs.haskell.org//srv/darcs/ghc
>
> On branch  : simd
>
>
> http://hackage.haskell.org/trac/ghc/changeset/08ba8f79c36d18598cd966b1f70c20575087fb32
>
> >---------------------------------------------------------------
>
> commit 08ba8f79c36d18598cd966b1f70c20575087fb32
> Author: Geoffrey Mainland <gmainlan at microsoft.com>
> Date:   Mon Oct 31 13:57:09 2011 +0000
>
>    Make sure the allocator aligns memory on a 16-byte boundary.
>
> >---------------------------------------------------------------
>
>  includes/rts/storage/Closures.h |    2 ++
>  rts/sm/Storage.c                |   13 +++++++++++++
>  2 files changed, 15 insertions(+), 0 deletions(-)
>
> diff --git a/includes/rts/storage/Closures.h
> b/includes/rts/storage/Closures.h
> index 5f4f035..c65d7da 100644
> --- a/includes/rts/storage/Closures.h
> +++ b/includes/rts/storage/Closures.h
> @@ -138,6 +138,8 @@ typedef struct StgBlockingQueue_ {
>  typedef struct {
>     StgHeader  header;
>     StgWord    bytes;
> +    StgWord    dummy1;
> +    StgWord    dummy2;
>     StgWord    payload[FLEXIBLE_ARRAY];
>  } StgArrWords;
>
> diff --git a/rts/sm/Storage.c b/rts/sm/Storage.c
> index 82e89a5..7d4526d 100644
> --- a/rts/sm/Storage.c
> +++ b/rts/sm/Storage.c
> @@ -601,12 +601,25 @@ move_STACK (StgStack *src, StgStack *dest)
>    Hp/HpLim).
>
>  --------------------------------------------------------------------------
> */
>
> +/*
> + * This is a terrible hack to guarantee that *all* allocated objects are
> aligned
> + * on a 4-word boundary so we can use aligned SSE instructions. We can't
> just
> + * align some objects, because when the GC runs and compresses the heap,
> they
> + * could end up non-aligned.
> + */
> +#define ALIGN_WORDS 4
> +#define ALIGN_MASK  (ALIGN_WORDS-1)
> +
> +#define ALIGN_ROUND_UP(n)  (((n)+ALIGN_WORDS-1) & ~ALIGN_MASK)
> +
>  StgPtr
>  allocate (Capability *cap, lnat n)
>  {
>     bdescr *bd;
>     StgPtr p;
>
> +    n = ALIGN_ROUND_UP(n);
> +
>     if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
>        lnat req_blocks =  (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
>
>
>
>
> _______________________________________________
> Cvs-ghc mailing list
> Cvs-ghc at haskell.org
> http://www.haskell.org/mailman/listinfo/cvs-ghc
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/cvs-ghc/attachments/20111101/2620312f/attachment.htm>