This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [gomp4] teach the tracer pass to ignore more blocks for OpenACC
- From: Nathan Sidwell <nathan_sidwell at mentor dot com>
- To: Cesar Philippidis <cesar at codesourcery dot com>, "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>
- Date: Wed, 26 Aug 2015 10:08:32 -0400
- Subject: Re: [gomp4] teach the tracer pass to ignore more blocks for OpenACC
- Authentication-results: sourceware.org; auth=none
- References: <55DDC5D1 dot 9060502 at codesourcery dot com>
On 08/26/15 09:57, Cesar Philippidis wrote:
I hit a problem in on one of my reduction test cases where the
GOACC_JOIN was getting cloned. Nvptx requires FORK and JOIN to be
single-entry, single-exit regions, or some form of thread divergence may
occur. When that happens, we cannot use the shfl instruction for
reductions or broadcasting (if the warp is divergent), and it may cause
problems with synchronization in general.
Nathan ran into a similar problem in one of the ssa passes when he added
support for predication in the nvptx backend. Part of his solution was
to add a gimple_call_internal_unique_p function to determine if internal
functions are safe to be cloned. This patch teaches the tracer to scan
each basic block for internal function calls using
gimple_call_internal_unique_p, and mark the blocks that contain certain
OpenACC internal functions calls as ignored. It is a shame that
gimple_statement_iterators do not play nicely with const_basic_block.
Is this patch ok for gomp-4_0-branch?
ok by me. (I idly wonder if tracer should be using the routine that
jump-threading has for scanning a block determining duplicability)
nathan
--
Nathan Sidwell - Director, Sourcery Services - Mentor Embedded