On Mon, Oct 26, 2015 at 10:36 AM, Richard Sandiford
<[email protected]> wrote:
> An upcoming patch adds a fold from hypot(x,x) to fabs(x)*sqrt(2).
> This is unusual in that it could trigger in the gimplifier but would
> require new SSA names to be created. This patch makes sure that we
> don't try to create new SSA names when we're not yet in SSA form.
>
> Tested on x86_64-linux-gnu, aarch64-linux-gnu and arm-linux-gnueabi.
> OK to install?
Please use
if (gimple_in_ssa_p (cfun))
res = make_ssa_name (type);
else
res = create_tmp_reg (type);
Ok with that change.
Thanks,
Richard.
> Thanks,
> Richard
>
>
> gcc/
> * gimple-match-head.c (maybe_push_res_to_seq): Don't attempt
> to create new SSA names if not in SSA form.
>
> diff --git a/gcc/gimple-match-head.c b/gcc/gimple-match-head.c
> index 8f72919..1345cf9 100644
> --- a/gcc/gimple-match-head.c
> +++ b/gcc/gimple-match-head.c
> @@ -331,7 +331,11 @@ maybe_push_res_to_seq (code_helper rcode, tree type,
> tree *ops,
> && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ops[2])))
> return NULL_TREE;
> if (!res)
> - res = make_ssa_name (type);
> + {
> + if (!gimple_in_ssa_p (cfun))
> + return NULL_TREE;
> + res = make_ssa_name (type);
> + }
> maybe_build_generic_op (rcode, type, &ops[0], ops[1], ops[2]);
> gimple *new_stmt = gimple_build_assign (res, rcode,
> ops[0], ops[1], ops[2]);
> @@ -361,7 +365,11 @@ maybe_push_res_to_seq (code_helper rcode, tree type,
> tree *ops,
> }
> gcc_assert (nargs != 0);
> if (!res)
> - res = make_ssa_name (type);
> + {
> + if (!gimple_in_ssa_p (cfun))
> + return NULL_TREE;
> + res = make_ssa_name (type);
> + }
> gimple *new_stmt = gimple_build_call (decl, nargs, ops[0], ops[1],
> ops[2]);
> gimple_call_set_lhs (new_stmt, res);
> gimple_seq_add_stmt_without_update (seq, new_stmt);
>
