On Thu, May 06, 2021 at 04:21:40PM +0200, Tobias Burnus wrote:
> * omp-low.c (lower_rec_simd_input_clauses): Set max_vf = 1 if
> a truth_value_p reduction variable is nonintegral.
> (lower_rec_input_clauses): Also handle SIMT part
> for complex/float recution with && and ||.
s/recution/reduction/
> --- a/gcc/omp-low.c
> +++ b/gcc/omp-low.c
> @@ -4389,14 +4389,28 @@ lower_rec_simd_input_clauses (tree new_var,
> omp_context *ctx,
> {
> for (tree c = gimple_omp_for_clauses (ctx->stmt); c;
> c = OMP_CLAUSE_CHAIN (c))
> - if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION
> - && OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
> - {
> - /* UDR reductions are not supported yet for SIMT, disable
> - SIMT. */
> - sctx->max_vf = 1;
> - break;
> + {
> + if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION)
> + continue;
> +
> + if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
> + {
> + /* UDR reductions are not supported yet for SIMT, disable
> + SIMT. */
> + sctx->max_vf = 1;
> + break;
> + }
> +
> + if (truth_value_p (OMP_CLAUSE_REDUCTION_CODE (c))
> + && !INTEGRAL_TYPE_P (TREE_TYPE (new_var)))
> + {
> + /* Doing boolean operations on non-boolean types is
> + for conformance only, it's not worth supporting this
> + for SIMT. */
This comment needs to be adjusted to talk about non-integral types.
> + sctx->max_vf = 1;
> + break;
> }
> + }
> }
> if (maybe_gt (sctx->max_vf, 1U))
> {
> @@ -6432,28 +6446,34 @@ lower_rec_input_clauses (tree clauses, gimple_seq
> *ilist, gimple_seq *dlist,
>
> gimplify_assign (unshare_expr (ivar), x, &llist[0]);
>
> - if (sctx.is_simt)
> - {
> - if (!simt_lane)
> - simt_lane = create_tmp_var (unsigned_type_node);
> - x = build_call_expr_internal_loc
> - (UNKNOWN_LOCATION, IFN_GOMP_SIMT_XCHG_BFLY,
> - TREE_TYPE (ivar), 2, ivar, simt_lane);
> - x = build2 (code, TREE_TYPE (ivar), ivar, x);
> - gimplify_assign (ivar, x, &llist[2]);
> - }
> tree ivar2 = ivar;
> tree ref2 = ref;
> + tree zero = NULL_TREE;
> if (is_fp_and_or)
> {
> - tree zero = build_zero_cst (TREE_TYPE (ivar));
> + zero = build_zero_cst (TREE_TYPE (ivar));
> ivar2 = fold_build2_loc (clause_loc, NE_EXPR,
> integer_type_node, ivar,
> zero);
> ref2 = fold_build2_loc (clause_loc, NE_EXPR,
> integer_type_node, ref, zero);
> }
> - x = build2 (code, TREE_TYPE (ref), ref2, ivar2);
> + if (sctx.is_simt)
> + {
> + if (!simt_lane)
> + simt_lane = create_tmp_var (unsigned_type_node);
> + x = build_call_expr_internal_loc
> + (UNKNOWN_LOCATION, IFN_GOMP_SIMT_XCHG_BFLY,
> + TREE_TYPE (ivar), 2, ivar, simt_lane);
> + if (is_fp_and_or)
> + x = fold_build2_loc (clause_loc, NE_EXPR,
> + integer_type_node, x, zero);
> + x = build2 (code, TREE_TYPE (ivar2), ivar2, x);
> + if (is_fp_and_or)
> + x = fold_convert (TREE_TYPE (ivar), x);
> + gimplify_assign (ivar, x, &llist[2]);
> + }
> + x = build2 (code, TREE_TYPE (ref2), ref2, ivar2);
> if (is_fp_and_or)
> x = fold_convert (TREE_TYPE (ref), x);
> ref = build_outer_var_ref (var, ctx);
Is this hunk still needed when the first hunk is in?
I mean, this is in code guarded with
is_simd && lower_rec_simd_input_clauses (...) and that function
will return false for if (known_eq (sctx->max_vf, 1U)) which the first hunk
ensures.
So sctx.is_simt && is_fp_and_or shouldn't be true in that code.
Jakub
