Re: [PATCH] Support AVX512F masked gather loads (PR tree-optimization/88464)

2018-12-13 Thread Jakub Jelinek
On Thu, Dec 13, 2018 at 05:09:17PM +0100, Richard Biener wrote: > Is there any chance you could look at replacing the i386 code to work with > the new IFN style used by Aarch64? Maybe for GCC 10, I'm afraid it is a lot of work mainly due to the weirdnesses of x86 mixed index vs. data sized types,

Re: [PATCH] Support AVX512F masked gather loads (PR tree-optimization/88464)

2018-12-13 Thread Richard Biener
On December 12, 2018 11:43:10 PM GMT+01:00, Jakub Jelinek wrote: >Hi! > >We support either the AVX2 gather loads (128-bit or 256-bit, using >masks in >vector registers), both conditional and unconditional, but AVX512F >gather >loads only unconditional. The problem was that tree-vect-stmts.c >did

[PATCH] Support AVX512F masked gather loads (PR tree-optimization/88464)

2018-12-12 Thread Jakub Jelinek
Hi! We support either the AVX2 gather loads (128-bit or 256-bit, using masks in vector registers), both conditional and unconditional, but AVX512F gather loads only unconditional. The problem was that tree-vect-stmts.c didn't have code to deal with the integral masktype argument the builtins have