Re: [PATCH] Remove more gimplified tree codes

[Devang Patel <dpatel at apple dot com>]

On Aug 6, 2004, at 8:02 AM, Paolo Bonzini wrote:

> The biggest cleaned up is in COND_EXPR; I left there basic support
> for COND_EXPR nodes since I heard that the vectorizer people wanted to
> reintroduce into GIMPLE very simple COND_EXPR nodes -- I did not ask
> if this addition was localized to the vectorizer or would have been
> propagated in the compiler.  It's only 60 lines of code after all.

We are preparing proposal about how to handle various vector 
operations. We are also discussing whether 'if-converted but not 
vectorized' code should be propagated in the compiler or not. thoughts? 
(In apple-ppc-branch, if-converted loop, that is not suitable for 
vectorization, is not propagated further. This is done be keep original 
version of loop around and selecting appropriate loop version based on 
vectorizer's decision.)

Richard H. has rejected the idea, offline, of using COND_EXPR nodes for 
vector operations in vectorized if-converted loop. So, if 'if-converted 
but not vectorized' is _not_ propagated further then we can remove 
COND_EXPR handling also.

-
Devang