On Tue, 2 Jan 2024 23:21:21 +0800
Chung-Lin Tang <clt...@pllab.cs.nthu.edu.tw> wrote:
> To Julian, there is a patch to the middle-end neutering, a hack
> actually, that detects SSA_NAMEs used in reduction array MEM_REFs,
> and avoids single->parallel copying (by moving those definitions
> before BUILT_IN_GOACC_SINGLE_COPY_START). This appears to work
> because reductions do their own initializing of the private copy.
It looks OK to me I think (bearing in mind your following paragraph, of
course!). I wonder though if maybe non-SSA (i.e. addressable) variables
need to be handled also, i.e. parts like this:
+ /* For accesses of variables used in array reductions, instead of
+ propagating the value for the main thread to all other worker threads
+ (which doesn't make sense as a reduction private var), move the defs
+ of such SSA_NAMEs to before the copy block and leave them alone (each
+ thread should access their own local copy). */
+ for (gimple_stmt_iterator i = gsi_after_labels (from); !gsi_end_p (i);)
+ {
+ gimple *stmt = gsi_stmt (i);
+ if (gimple_assign_single_p (stmt)
+ && def_escapes_block->contains (gimple_assign_lhs (stmt))
+ && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME)
are only handling SSA-converted variables. But maybe that's OK?
> As we discussed in our internal calls, the real proper way is to
> create the private array in a more appropriate stage, but that is too
> long a shot for now. The changes here are needed at least for some
> -O0 cases (when under optimization, propagation of the private
> copies' local address eliminate the SSA_NAME and things actually just
> work in that case). So please bear with this hack.
HTH,
Julian
