Re: [rfc rs6000] troubles with gimple folding for vec_sel

2018-10-24 Thread Segher Boessenkool
On Wed, Oct 24, 2018 at 02:27:07PM +0200, Richard Biener wrote: > If vec_sel is really a bitwise merge then there's no choice but using > BIT_{AND,IOR,NOT}_EXPR for open-coding it. I suppose the > original builtin expanded to an UNSPEC because there's nothing > in RTL besides bitwise operations de

Re: [rfc rs6000] troubles with gimple folding for vec_sel

2018-10-24 Thread Richard Biener
On Tue, Oct 23, 2018 at 11:47 PM Will Schmidt wrote: > > Hi all, > I've been attempting to get early gimple-folding to work with the > vec_sel intrinsic for powerpc, and I've run into a snag or two such that > I'm not sure how to best proceed. Code snippet is below, followed by a > description of

[rfc rs6000] troubles with gimple folding for vec_sel

2018-10-23 Thread Will Schmidt
Hi all, I've been attempting to get early gimple-folding to work with the vec_sel intrinsic for powerpc, and I've run into a snag or two such that I'm not sure how to best proceed. Code snippet is below, followed by a description of the issues as I interpret them below. Apologies for the ramble,