> Well, the current code assumes that the replacements are gimple
> register types and, perhaps more importantly, it assumes there are no
> to-be-replaced pieces within to-be-replaced pieces. If we were to put
> a structure outside of a frame structure and scalarize some field
> within it at the s
On Fri, Sep 21, 2012 at 01:32:25PM +0200, Richard Guenther wrote:
...
> Yes, most of the SRA heuristic games make it complicated and ugly,
> especially as it is isn't clearly separate analysis / decision / transform
> phases. TLC welcome ;)
>
While I agree that the heuristics of SRA is surpris
Hi,
On Fri, Sep 21, 2012 at 12:48:16PM +0200, Eric Botcazou wrote:
>
...
> > As of doing it in SRA what I'd do there is special-case FRAME for both
> > candidate consideration (so you get around the addressable issue)
> > and replacement generation.
>
> OK, but you need to be able to split the
On Fri, Sep 21, 2012 at 12:48 PM, Eric Botcazou wrote:
>> You rely on being able to see all FRAME accesses as component refs,
>> thus nothing transforms them into just MEM[&FRAME, offset]. That's of
>> course something that can be easily "broken" by means of doing
>> some pointer arithmetic like
> You rely on being able to see all FRAME accesses as component refs,
> thus nothing transforms them into just MEM[&FRAME, offset]. That's of
> course something that can be easily "broken" by means of doing
> some pointer arithmetic like (untested, but you get the idea)
>
> foo()
> {
> int c[32
On Thu, Sep 20, 2012 at 12:56 PM, Eric Botcazou wrote:
>> I really don't like this to be done outside of SRA (and it is written in a
>> non-MEM_REF way).
>
> Could you elaborate on the latter point? If it can be improved, even in its
> current form...
You rely on being able to see all FRAME acce
> I really don't like this to be done outside of SRA (and it is written in a
> non-MEM_REF way).
Could you elaborate on the latter point? If it can be improved, even in its
current form...
> For the testcase in question we scalarize back
> 'i' in SRA (other scalars are optimized away already, b
On Wed, Sep 19, 2012 at 01:36:50PM +0200, Richard Guenther wrote:
> On Wed, Sep 19, 2012 at 12:58 PM, Eric Botcazou wrote:
> I really don't like this to be done outside of SRA (and it is written in a
> non-MEM_REF way). For the testcase in question we scalarize back
> 'i' in SRA (other scalars ar
On Wed, Sep 19, 2012 at 12:58 PM, Eric Botcazou wrote:
> Hi,
>
> this transformation has been in our tree for a couple of years and was
> originally developed for SPARKSkein (http://www.skein-hash.info/node/48),
> which is the implementation in SPARK (subset of Ada) of the Skein algorithm.
>
> Whe
Hi,
this transformation has been in our tree for a couple of years and was
originally developed for SPARKSkein (http://www.skein-hash.info/node/48),
which is the implementation in SPARK (subset of Ada) of the Skein algorithm.
When nested functions access local variables of their parent, the com
10 matches
Mail list logo
